Directory
U.S.-led models, software, datasets, and evaluation tools with publicly available materials. Each record says what is public, under which license, and who maintains it. Publicly downloadable does not always mean unrestricted.
178 of 178 records shown
MLCommons (AI Risk & Reliability working group)
AILuminate is MLCommons' family of safety and security benchmarks for generative AI systems, organized around 12 hazard categories. The v1.0 safety benchmark (December 2024) tests general-purpose chat systems with more than 24,000 human-written prompts and grades responses with an ensemble of evaluator models; a jailbreak benchmark measures how safety degrades under attack.
ServiceNow (SLAM lab)
Apriel is ServiceNow's series of small open-weight language models, built by its SLAM (ServiceNow Language Models) lab. Apriel-5B was the first release. It was followed by 15B-parameter reasoning models: Apriel-Nemotron-15b-Thinker, the multimodal Apriel-1.5-15b-Thinker, and Apriel-1.6-15b-Thinker (December 2025). In April 2026 ServiceNow released Super Apriel, a 15B model derived from Apriel 1.6 whose layers can switch between attention and linear-time mixers at inference time.
ServiceNow (SLAM lab) · Apriel family
A multimodal reasoning model of about 15B parameters that accepts text and images and writes its reasoning steps before a marked final answer. ServiceNow released it in December 2025 as an update to Apriel-1.5-15b-Thinker, with further continual pretraining, supervised fine-tuning, and reinforcement learning aimed partly at shorter reasoning.
ARC Prize Foundation
ARC-AGI is a series of benchmarks that measure how efficiently a system acquires new skills on unfamiliar tasks, using tasks that rely only on basic "core knowledge" priors and are easy for people. ARC-AGI-1 (introduced in 2019) and ARC-AGI-2 (March 2025) are grid puzzles in which the solver infers a transformation from a few input-output examples; ARC-AGI-3 (March 2026) consists of interactive, turn-based game environments in which agents must explore and infer goals without explicit instructions. Public tasks are released for each version, while semi-private and private evaluation sets used for official scoring are held back.
Microsoft
AutoGen is Microsoft's open-source Python framework for building multi-agent AI applications that act autonomously or alongside people. It comprises an event-driven Core, the higher-level AgentChat API, Extensions for external services such as MCP servers and code execution, and the no-code AutoGen Studio. The project is now in maintenance mode, and Microsoft directs new users to its successor, Microsoft Agent Framework.
Gorilla project, UC Berkeley
The Berkeley Function Calling Leaderboard (BFCL) is a benchmark and live leaderboard from UC Berkeley's Gorilla project for evaluating how well language models call functions and tools. It checks serial and parallel function calls in several programming languages using abstract syntax tree (AST) evaluation. Later versions add multi-turn and multi-step scenarios and, in V4, agentic tasks such as web search, memory management, and format sensitivity.
Chroma
Chroma is an open-source database for AI applications that stores documents, embeddings, and metadata and supports vector, full-text, regex, and metadata search. It can run in memory inside a program, persist to local disk, or run as a server, with client libraries for Python, JavaScript/TypeScript, and Rust.
Amazon (Amazon Web Services)
Chronos is Amazon's family of pretrained time-series forecasting models. It includes the original T5-based Chronos models (March 2024, 8M to 710M parameters), the patch-based Chronos-Bolt models (November 2024), and Chronos-2 (October 2025), a 120M-parameter encoder-only model that adds multivariate and covariate-informed forecasting.
Amazon (Amazon Web Services) · Chronos (time series models) family
Chronos-2 is a 120M-parameter, encoder-only time-series foundation model from Amazon, released on October 20, 2025. Inspired by the T5 encoder, it produces multi-step quantile forecasts and uses group attention to learn in context across related series and covariates, supporting univariate, multivariate, and covariate-informed forecasting with contexts up to 8,192 steps and horizons up to 1,024 steps.
OpenAI
CLIP (Contrastive Language-Image Pre-Training) is an OpenAI family of paired image and text encoders trained with a contrastive objective on 400 million image-text pairs, which lets the model classify images zero-shot from natural-language labels. OpenAI released the models in stages from January 2021 to April 2022, from ResNet-50 and ViT-B/32 up to ViT-L/14 and ViT-L/14 at 336-pixel resolution.
OpenAI · CLIP family
CLIP ViT-L/14 pairs a ViT-L/14 Vision Transformer image encoder with a masked self-attention Transformer text encoder, trained contrastively to match images with their captions. OpenAI released it in January 2022; a variant further trained at 336-pixel resolution (ViT-L/14@336px) followed in April 2022.
Deep Cogito
Cogito is Deep Cogito's series of open-weight language models; the v2 and v2.1 models are hybrid reasoning models that can answer directly or reason first. Cogito v1 Preview (April 2025) covered 3B to 70B models on Llama and Qwen bases; Cogito v2 Preview (July 2025) added 70B dense, 109B MoE, 405B dense, and 671B MoE models on Llama and DeepSeek bases; and Cogito v2.1 (November 2025) is a 671B MoE model with 37B active parameters on a DeepSeek base.
Deep Cogito · Cogito family
A 405B-parameter dense hybrid reasoning model that Deep Cogito post-trained from Meta's Llama 3.1 405B, released in July 2025 as one of four Cogito v2 Preview models. It can answer directly or reason first, supports a 128k-token context, and was trained in over 30 languages using Deep Cogito's Iterated Distillation and Amplification approach.
Deep Cogito · Cogito family
A 671B-parameter mixture-of-experts model with 37B active parameters that Deep Cogito post-trained in-house from the DeepSeek-V3 base model. It is a hybrid reasoning model that can answer directly or reason first, supports a 128k-token context, and was trained in over 30 languages; Deep Cogito says process supervision of reasoning chains lets it reach answers with shorter reasoning.
Common Crawl Foundation
An archive of web crawl data collected by Common Crawl's crawlers since 2008, released as periodic crawls. Each crawl is published as WARC files (raw HTTP responses, requests, and metadata), WAT files (computed metadata as JSON), and WET files (extracted plain text), together with URL indexes.
PyTorch Foundation (hosted project), DeepSpeed Technical Steering Committee (project committers)
DeepSpeed is an open-source PyTorch library for distributed training and inference of large models; its features include the ZeRO memory optimizations, 3D parallelism, Ulysses sequence parallelism, and mixture-of-experts support. Microsoft contributed it to LF AI & Data as an incubation project in February 2025, and in May 2025 it became a hosted project of the PyTorch Foundation, which names Microsoft as the contributor.
Hugging Face
Diffusers is Hugging Face's open-source PyTorch library for pretrained diffusion models that generate images, video, and audio. It is organized around the DiffusionPipeline API, with interchangeable noise schedulers and model components, and supports adapters such as LoRA. Version 0.40.0 (August 2026) removed JAX/Flax support.
Meta (FAIR)
DINOv3 is Meta's third generation of self-supervised vision backbones, announced in August 2025. The suite includes a 6.7B-parameter ViT-7B/16, distilled Vision Transformers from 21M to 840M parameters, ConvNeXt models, and two backbones trained on satellite imagery, plus task heads for classification, depth, detection, and segmentation. The web-image models were trained on LVD-1689M, a curated set of about 1.7 billion images.
Meta (FAIR) · DINOv3 family
The largest DINOv3 backbone: a 6.7-billion-parameter Vision Transformer with 16-pixel patches, trained by Meta with self-supervised learning (a DINO self-distillation loss and an iBOT masked-image-modeling loss) on LVD-1689M. Training ran in three stages: pretraining, Gram anchoring, and high-resolution adaptation.
LF AI & Data Foundation (hosted project), IBM Research (originating team)
Docling is an open-source Python toolkit that converts documents such as PDF, DOCX, PPTX, XLSX, HTML, images, audio, and other formats into a structured DoclingDocument that can be exported as Markdown, HTML, or JSON for generative AI applications. It includes PDF layout, reading-order, and table-structure analysis, OCR, and optional vision-language model pipelines. The project began at IBM Research in Zurich and has been hosted by the LF AI & Data Foundation since 2025.
Ai2 (Allen Institute for AI)
Dolma is Ai2's family of English pretraining corpora built from web pages, academic publications, code, and encyclopedic text, used to train the OLMo models. The first release (v1, August 2023) was updated through v1.7 (April 2024); Dolma 3, used for Olmo 3, consists of a pool of about 9.3 trillion tokens and curated pretraining, mid-training, and long-context mixes. A Dolma 3.5 pool with additional sources and filtering has since been published.
Stanford NLP Group
DSPy is an open-source Python framework for building language-model programs from compositional Python modules instead of hand-written prompts. It includes optimizer algorithms that tune the prompts and weights of those programs. The project started at the Stanford NLP group.
Biohub (Chan Zuckerberg Biohub, Inc.), EvolutionaryScale (original developer; now part of Biohub)
Protein language models developed by EvolutionaryScale and, since its team joined Biohub in November 2025, released by Biohub. The line includes ESM3 (June 2024), a generative model that reasons over protein sequence, structure, and function, with a largest version of 98B parameters and a small open model; ESM C (ESM Cambrian, December 2024), representation models at 300M, 600M, and 6B parameters; and, from May 2026, ESMFold2 structure prediction models and sparse autoencoders built on ESM C.
Biohub (Chan Zuckerberg Biohub, Inc.), EvolutionaryScale (original developer; now part of Biohub) · ESM (EvolutionaryScale and Biohub protein models) family
The openly released small ESM3 model, which EvolutionaryScale describes as a 1.4B-parameter open model. ESM3 is a generative masked language model that takes partial prompts across protein sequence, structure, and function tracks and predicts all three. EvolutionaryScale released it in June 2024; Biohub now distributes it under the MIT license.
Biohub (Chan Zuckerberg Biohub, Inc.), EvolutionaryScale (original developer; now part of Biohub) · ESM (EvolutionaryScale and Biohub protein models) family
The largest ESM C (ESM Cambrian) protein language model: 6 billion parameters, 80 transformer layers, and 2.37e23 training FLOPs. EvolutionaryScale introduced ESM C in December 2024; Biohub's May 2026 release of ESMC, ESMFold2, and ESM Atlas included the 6B weights, and ESMFold2 is trained on top of a frozen ESMC 6B.
Essential AI
Essential-Web v1.0 is a web text dataset of about 24 trillion tokens in 23.6 billion documents, built by Essential AI from 101 Common Crawl snapshots. Every document carries metadata from a 12-category taxonomy covering subject (Free Decimal Correspondence, a Dewey Decimal-inspired scheme), page type, reasoning depth, education level, and quality, so subsets can be selected with metadata filters.
Arc Institute
Evo 2 is a family of DNA language models from Arc Institute, developed with NVIDIA and collaborators at institutions including Stanford University, UCSF, UC Berkeley, Goodfire, and the University of Washington. The models use the StripedHyena 2 architecture, work at single-nucleotide resolution with context of up to 1 million base pairs, and were trained on the OpenGenome2 dataset. Checkpoints range from 1B to 40B parameters; a 20B checkpoint derived from the 40B model was added in February 2026.
Arc Institute · Evo 2 family
A 20-billion-parameter Evo 2 checkpoint with 1 million base-pair context, released in February 2026. Arc created it from Evo 2 40B without additional training by removing layers that logit-lens analysis showed could be dropped with little effect on loss, while keeping the unembedding layer.
Arc Institute · Evo 2 family
The 40-billion-parameter Evo 2 checkpoint, with 50 layers and a context length of up to 1 million base pairs. It was trained autoregressively on OpenGenome2 on 2,048 GPUs, per the Savanna README; a separate evo2_40b_base checkpoint trained at 8,192-token context is also published.
Apple
FastVLM is a family of vision-language models from Apple, described in a CVPR 2025 paper, built around FastViTHD, a hybrid convolutional-transformer vision encoder that outputs fewer visual tokens to reduce encoding time for high-resolution images. Apple released 0.5B, 1.5B, and 7B variants that pair the encoder with Qwen2 language models, with PyTorch checkpoints and versions exported for Apple silicon.
Apple · FastVLM family
FastVLM 7B is the 7B variant of Apple's FastVLM release. It combines Apple's FastViTHD vision encoder with the Qwen2-7B language model in a LLaVA-style architecture and accepts an image plus a text prompt to generate text.
Hugging Face (FineData, HuggingFaceFW)
FineWeb is Hugging Face's English web-text pretraining dataset, built by extracting, filtering, and deduplicating pages from Common Crawl snapshots dating back to 2013. The card describes it as more than 18.5 trillion tokens (GPT-2 tokenizer), up from about 15 trillion at first release in April 2024; version 1.4.0 (July 2025) added the Common Crawl snapshots from January to June 2025. Hugging Face also published FineWeb-Edu, an educational subset.
Google DeepMind
Gemma is a family of open-weight models credited to Google DeepMind. Google's release page lists general-purpose generations including Gemma 3 (2025) and Gemma 4 (2026), along with specialized variants such as EmbeddingGemma, MedGemma, ShieldGemma, and FunctionGemma.
Google DeepMind · Gemma family
Gemma 4 12B Unified is an 11.95-billion-parameter Gemma 4 model with an encoder-free design: image patches and audio waveforms are projected directly into the language model instead of passing through separate encoders. It accepts text, image, audio, and video (as frames) input and generates text, with a 256K-token context window. Google's release page dates its release to June 3, 2026, after the other Gemma 4 sizes.
Google DeepMind · Gemma family
Gemma 4 26B A4B is the mixture-of-experts model in the Gemma 4 generation, with 25.2 billion total and 3.8 billion active parameters (8 of 128 experts active, plus one shared expert). It accepts text and image input, can process video as frames, and generates text, with a 256K-token context window. Google's release page dates the initial Gemma 4 release, which included this size, to March 31, 2026.
Google DeepMind · Gemma family
Gemma 4 31B is the dense model in the Gemma 4 generation, with 30.7 billion parameters. It accepts text and image input, can process video as frames, and generates text, with a 256K-token context window. It is published as pre-trained and instruction-tuned checkpoints; Google's release page dates the initial Gemma 4 release, which included this size, to March 31, 2026.
EleutherAI
GPT-NeoX is EleutherAI's library for training large autoregressive language models on GPUs. It builds on NVIDIA's Megatron-LM and on DeepSpeed, and supports distributed training with ZeRO and 3D parallelism, several positional-embedding and attention options, mixture of experts, and predefined configurations for architectures such as Pythia and Llama. EleutherAI used it to train GPT-NeoX-20B and the Pythia suite.
OpenAI
gpt-oss is OpenAI's series of open-weight, text-only reasoning models, released in August 2025 in two sizes: gpt-oss-120b and gpt-oss-20b. Both are mixture-of-experts transformers trained for OpenAI's harmony response format. OpenAI later published gpt-oss-safeguard models fine-tuned from them for safety classification.
OpenAI · gpt-oss family
gpt-oss-120b is the larger gpt-oss model: a 36-layer mixture-of-experts transformer with 116.8B total and about 5.1B active parameters per token, using 128 experts with the top 4 selected per token. It is text-only, supports context up to 131,072 tokens, and ships with its MoE weights quantized to MXFP4.
OpenAI · gpt-oss family
gpt-oss-20b is the smaller gpt-oss model: a 24-layer mixture-of-experts transformer with 20.9B total and about 3.6B active parameters per token, using 32 experts with the top 4 selected per token. It is text-only, supports context up to 131,072 tokens, and ships with its MoE weights quantized to MXFP4.
Nomic AI
GPT4All is Nomic's open-source desktop application for running large language models locally on Windows, macOS, and Linux. It includes LocalDocs for chatting with local files, a Python SDK, and an API server, and its Python client wraps llama.cpp.
IBM (Granite Team) · IBM Granite family
Granite 4.2 30B is a 30-billion-parameter dense decoder-only language model from IBM with built-in reasoning, selectable thinking, non-thinking, and low-effort modes, and tool calling. It supports a 128K-token context natively; the model card documents extension to 512K. Per the model card, the 30B model received a second supervised fine-tuning phase with up-sampled agentic data.
IBM (Granite Team) · IBM Granite family
Granite 4.2 3B is a 3-billion-parameter dense decoder-only language model from IBM with built-in reasoning, selectable thinking, non-thinking, and low-effort modes, and tool calling. It supports a 128K-token context natively; the model card documents extension to 512K.
IBM (Granite Team) · IBM Granite family
Granite 4.2 8B is an 8-billion-parameter dense decoder-only language model from IBM with built-in reasoning, selectable thinking, non-thinking, and low-effort modes, and tool calling. It supports a 128K-token context natively; the model card documents extension to 512K.
xAI (SpaceXAI)
Grok is xAI's family of large language models, first launched in November 2023 and offered mainly through the Grok assistant and the xAI API. xAI has published weights for two earlier models: Grok-1 in March 2024 and Grok 2 in August 2025. SpaceX's prospectus describes current Grok models as proprietary.
xAI (SpaceXAI) · Grok family
Grok 2 is a post-trained language model that xAI trained and used in 2024; its weights, about 500 GB in 42 files, were published on Hugging Face in August 2025. Press coverage at the time reported the release as Grok 2.5.
xAI (SpaceXAI) · Grok family
Grok-1 is a 314-billion-parameter mixture-of-experts language model (8 experts, 2 used per token) that xAI trained from scratch. The published checkpoint is the raw base model from pretraining that ended in October 2023 and is not fine-tuned for dialogue or other applications.
Stanford Center for Research on Foundation Models (CRFM)
HELM is an open-source Python framework from Stanford's Center for Research on Foundation Models for evaluating foundation models, including language and multimodal models. It provides benchmarks in a standardized format, one interface to models from several providers, metrics beyond accuracy such as efficiency, bias, and toxicity, and a web UI and leaderboards for inspecting results. HELM entered maintenance mode on June 1, 2026.
Nous Research
Hermes is Nous Research's series of post-trained language models, each built on a base model released by another developer. Releases listed by Nous include Nous-Hermes Llama 2 13B (July 2023), Hermes 2 models on Mistral, Mixtral, Yi, and Llama 3 bases (2023 to 2024), Hermes 3 on Llama 3.1 (August 2024), Hermes 4 at 405B, 70B, and 14B (August 2025), and Hermes 4.3 36B (December 2025). Hermes 4 and 4.3 are hybrid reasoning models that can answer directly or reason in think tags first.
Nous Research · Hermes family
A 70B-parameter hybrid reasoning model that Nous Research post-trained from Meta's Llama 3.1 70B base model, released in August 2025 with Hermes 4 405B and 14B. Nous trained it on a newly synthesized dataset of about 5 million reasoning and non-reasoning samples, and it can either answer directly or reason inside think tags first.
Nous Research · Hermes family
A 36B-parameter hybrid reasoning model that Nous Research post-trained from ByteDance Seed's Seed-OSS-36B-Base. It was Nous's first Hermes model trained in a decentralized way over the internet on the Psyche network, with context extended up to 512K tokens. Nous also released a centrally trained version of the same model as a research comparison.
OpenAI
HumanEval is a code-generation benchmark released by OpenAI with the Codex paper in July 2021. It contains 164 hand-written Python programming problems, each with a function signature, docstring, reference solution, and unit tests, and measures functional correctness of generated code with the pass@k metric. The GitHub repository provides the problem file and the evaluation harness.
Center for AI Safety, Scale AI
A multimodal benchmark of 2,500 expert-written, closed-ended academic questions across more than a hundred subjects, including mathematics, the humanities, and the natural sciences, in multiple-choice and short-answer formats suited to automated grading. It was organized by teams at the Center for AI Safety and Scale AI with questions from nearly 1,000 subject-matter contributors, and was published in Nature in January 2026.
IBM
Granite is IBM's family of models. IBM lists language, speech, vision, guardian (risk detection), embedding, and time-series lines. The most recent language-model generation on IBM's Granite page is Granite 4.2: dense reasoning models in 3B, 8B, and 30B sizes, released in August 2026 and post-trained from Granite 4.1 base models.
Thinking Machines Lab · Inkling family
Inkling is a 66-layer decoder-only mixture-of-experts transformer with 975B total and 41B active parameters, routing each token to 6 of 256 experts plus 2 shared experts. It accepts text, image, and audio input, outputs text, and supports a context window of up to 1M tokens. Thinking Machines Lab says it trained the model from scratch on 45 trillion tokens of text, images, audio, and video.
Thinking Machines Lab
Inkling is Thinking Machines Lab's family of open-weight multimodal mixture-of-experts models that accept text, image, and audio input and produce text. The first release, Inkling (975B total and 41B active parameters), was published on July 15, 2026, followed by Inkling-Small (276B total and 12B active) on July 30, 2026.
Thinking Machines Lab · Inkling family
Inkling-Small is a 42-layer decoder-only mixture-of-experts transformer with 276B total and 12B active parameters, routing each token to 6 of 256 experts plus 2 shared experts. Like Inkling, it accepts text, image, and audio input, outputs text, and supports a context window of up to 1M tokens.
Prime Intellect
INTELLECT is Prime Intellect's series of open-weight models. INTELLECT-1 is a 10B model trained from scratch on 1 trillion tokens with compute from distributed community contributors; INTELLECT-2 is a 32B model produced by a distributed reinforcement-learning run on QwQ-32B; INTELLECT-3 (November 2025) is a 106B mixture-of-experts model (12B active) post-trained from GLM-4.5-Air-Base with supervised fine-tuning and large-scale RL; and INTELLECT-3.1 continues INTELLECT-3 with further RL.
Prime Intellect · INTELLECT family
A 106B-parameter mixture-of-experts reasoning model with 12B active parameters, post-trained by Prime Intellect from Z.ai's GLM-4.5-Air-Base. Training used two supervised fine-tuning stages (general reasoning, then agentic) followed by large-scale reinforcement learning on math, code, science, logic, deep-research, and software-engineering environments, run with prime-rl on a 512-GPU H200 cluster over about two months.
Prime Intellect · INTELLECT family
A 106B-parameter mixture-of-experts reasoning model (12B active) that continues the training of INTELLECT-3 with additional reinforcement learning on math, coding, software-engineering, and agentic tasks, using prime-rl and environments built with the verifiers library.
JAX core team (Google open-source project)
JAX is a Python library for array computation and program transformation, including automatic differentiation, just-in-time compilation, and vectorization of NumPy-style code. It compiles programs with the XLA compiler to run on CPUs, GPUs, TPUs, and other accelerators.
Keras team at Google
Keras is an open-source deep learning API. Keras 3 is multi-backend: models can run on JAX, TensorFlow, or PyTorch, and an OpenVINO backend supports inference only. The keras.io site also documents KerasHub, which provides Keras 3 implementations of model architectures paired with pretrained checkpoints.
Poolside
Laguna is Poolside's family of mixture-of-experts language models for agentic coding. Poolside's release notes and announcements list Laguna XS.2 and Laguna M.1 (April 2026), Laguna XS 2.1 (announced 2 July 2026), and Laguna S 2.1 (21 July 2026), with sizes from 33B total (3B active) to 225B total (23B active) parameters.
Poolside · Laguna family
Laguna S 2.1 is a mixture-of-experts model with 118B total and about 8B active parameters per token, built for agentic coding. It has 48 layers mixing global and sliding-window attention, 256 routed experts plus one shared expert, and a 1,048,576-token context window, with reasoning that can be switched on or off per request.
Poolside · Laguna family
Laguna XS 2.1 is a mixture-of-experts model with 33B total and 3B active parameters per token, built for agentic coding on a local machine. Poolside describes it as an upgraded version of Laguna XS.2 with the same architecture. It supports interleaved reasoning between tool calls and has a 262,144-token context window.
LangChain Inc.
LangChain is an open-source Python framework for building agents and LLM applications. It gives a standard interface across model providers and many third-party integrations. Its agent API is built on LangGraph, and a JavaScript/TypeScript version, LangChain.js, lives in a separate repository.
LangChain Inc.
LangGraph is an open-source, low-level orchestration framework and runtime for long-running, stateful agents. It is built by LangChain Inc. but can be used without the LangChain library, and a JavaScript/TypeScript version, LangGraph.js, lives in a separate repository. The README credits Pregel and Apache Beam as inspirations.
Liquid AI
LFM is Liquid AI's series of hybrid language and multimodal models for on-device use, built from gated short-convolution blocks and a small number of grouped-query attention blocks. LFM2 (from July 2025) introduced open-weight dense models from 350M parameters upward and later an 8.3B mixture-of-experts model. LFM2.5 (from January 2026) extended pretraining and scaled up reinforcement-learning post-training; 2026 releases include LFM2.5-8B-A1B (May), LFM2.5-2.6B (August), and LFM2.5-VL-3B (August), alongside vision, audio, and encoder variants.
Liquid AI · LFM (Liquid Foundation Models) family
A 2.69B-parameter text-only hybrid model with 30 layers (22 double-gated short-convolution blocks and 8 grouped-query attention blocks), a 131,072-token context window, and support for 16 languages. Liquid AI pretrained it on about 34 trillion tokens and post-trained it for agentic use; it always reasons before answering.
Liquid AI · LFM (Liquid Foundation Models) family
A text-only mixture-of-experts model with 8.3B total and 1.5B active parameters, built from 18 double-gated convolution blocks and 6 grouped-query attention blocks. Liquid AI pretrained it on 38 trillion tokens, extended its context to 128,000 tokens, and tuned it for reasoning; it writes a chain of thought before its final answer. It succeeds LFM2-8B-A1B.
Liger Kernel is an open-source collection of Triton GPU kernels for training large language models, developed at LinkedIn. It provides Hugging Face-compatible implementations of layers such as RMSNorm, RoPE, SwiGLU, and cross-entropy (including a fused linear cross-entropy), plus memory-efficient losses for post-training methods such as DPO, ORPO, and KTO.
Meta
Llama is Meta's family of large language models. Meta's Llama models repository lists generations from Llama 2 (July 2023) through Llama 4 (April 2025), each with its own license file and acceptable use policy. Llama 4 is distributed under the custom Llama 4 Community License Agreement rather than a standard open-source license.
Ai2 (Allen Institute for AI) · Tülu family
An 8B instruction-following model from Ai2's Tülu 3 line, post-trained from Meta's Llama 3.1 8B base model through supervised fine-tuning and DPO, then reinforcement learning with verifiable rewards. Version 3.1 differs from the original Tülu 3 8B only in the final RL stage, which switched from PPO to GRPO (without a reward model) with further hyperparameter tuning.
Meta · Llama family
Llama 4 Maverick is a natively multimodal mixture-of-experts model from Meta with 17 billion active and 400 billion total parameters across 128 experts. It accepts multilingual text and images and produces text and code; the model card lists a 1M-token context length and an August 2024 knowledge cutoff. Pretrained and instruction-tuned versions were released.
Meta · Llama family
Llama 4 Scout is a natively multimodal mixture-of-experts model from Meta with 17 billion active and 109 billion total parameters across 16 experts. It accepts multilingual text and images and produces text and code; the model card lists a 10M-token context length and an August 2024 knowledge cutoff. Pretrained and instruction-tuned versions were released.
LlamaIndex, Inc.
LlamaIndex is an open-source Python framework for building LLM agents, event-driven workflows, and retrieval-augmented generation (RAG) over private data. It is split into a core package and hundreds of integration packages for models, embeddings, and vector stores; the company also offers a TypeScript SDK. The README says the company now focuses mainly on its LlamaParse platform but keeps the framework available as an open toolkit.
EleutherAI
An open-source framework from EleutherAI for evaluating language models on many benchmark tasks through one interface. It supports local models (for example through Hugging Face Transformers or vLLM) and hosted model APIs, and its README lists more than 60 standard academic benchmarks with hundreds of subtasks and variants.
Modular
MAX is Modular's framework for serving and building AI models. It serves models through an OpenAI-compatible API, offers PyTorch-like Python APIs for writing model pipelines, and includes a GPU kernel library for NVIDIA, AMD, and Apple hardware. Its Python API, model pipelines, and GPU kernels are open source; the MAX package as a whole is distributed under the Modular Community License.
Google (AI Hypercomputer GitHub organization)
MaxText is an open-source LLM library and reference implementation written in Python and JAX for training on Google Cloud TPUs and GPUs. It provides implementations of Gemma, Llama, DeepSeek, Qwen, Mistral, and other open model architectures, with support for pre-training and for post-training methods such as supervised fine-tuning and reinforcement learning.
Google (Health AI Developer Foundations)
MedGemma is a collection of Gemma 3 variants trained by Google on medical text and images, published as part of Health AI Developer Foundations (HAI-DEF). MedGemma 1 was released in May 2025 (4B multimodal and 27B text-only) with a 27B multimodal variant in July 2025, and MedGemma 1.5 4B followed in January 2026.
Google (Health AI Developer Foundations) · MedGemma family
MedGemma 1.5 4B is a multimodal, instruction-tuned update to MedGemma 1 4B, released by Google on January 13, 2026. It accepts text and images and produces text, adding support for CT and MRI volumes, whole-slide histopathology, longitudinal chest X-rays, anatomical localization, lab-report extraction, and text-based EHR interpretation.
Google (Health AI Developer Foundations) · MedGemma family
MedGemma 27B multimodal is the 27B multimodal variant of MedGemma 1, released by Google on July 9, 2025. It accepts text and images and produces text, and unlike the 27B text-only variant it was also trained on medical images and FHIR-based electronic health record data. It is published only as an instruction-tuned model.
NVIDIA
NVIDIA's Megatron-LM repository contains two components: Megatron Core, a library of GPU-optimized building blocks for training transformer models at scale (tensor, pipeline, data, expert, and context parallelism; FP16, BF16, FP8, and FP4 mixed precision), and Megatron-LM, a reference training setup with pre-configured scripts built on Megatron Core. The README says Megatron Core development moved to GitHub in December 2025, with all development and CI now happening in the open.
MLflow Project, a Series of LF Projects, LLC (Linux Foundation), Databricks (creator)
MLflow is an open-source platform for managing machine learning and generative AI development. For LLM applications and agents it provides tracing and observability (with OpenTelemetry integration), evaluation, prompt management and optimization, and an AI Gateway; for model training it provides experiment tracking, model evaluation, a model registry, and deployment tools. Databricks created MLflow, which joined the Linux Foundation in June 2020.
MLCommons
MLPerf is MLCommons' family of system performance benchmarks. Its suites cover training (time to train a model to a target quality), inference in datacenter, edge, mobile, and tiny settings, and also client PCs, storage, automotive, HPC training, and inference endpoints. Results are submitted in rounds, reviewed by the submitting organizations, and published together by MLCommons.
Apple (machine learning research)
MLX is an array framework for machine learning on Apple silicon from Apple machine learning research. It has a NumPy-like Python API plus C++, C, and Swift APIs, composable function transformations, lazy computation, and a unified memory model that shares arrays between the CPU and GPU.
Genmo
Mochi is Genmo's line of text-to-video diffusion models. Its only open-weight release so far is Mochi 1 preview (October 2024), a 10B-parameter model published with its video VAE under Apache 2.0. Genmo's hosted playground refers to a Mochi 1.1 model, for which this catalog found no open-weight release.
Genmo · Mochi family
A 10B-parameter text-to-video diffusion model built on Genmo's Asymmetric Diffusion Transformer (AsymmDiT) and released as a research preview in October 2024. It is paired with AsymmVAE, a 362M-parameter video autoencoder released alongside it, and encodes prompts with a single T5-XXL model. The initial release generates 480p video.
Modular
Mojo is a programming language developed by Modular that pairs Python-style syntax with systems programming and metaprogramming features, for writing code that runs on CPUs and GPUs. Its compiler, standard library, and tooling are published in the Mojo directory of Modular's modular repository; Mojo 1.0 was released in August 2026.
Ai2 (Allen Institute for AI)
Molmo is Ai2's family of open vision-language models. The first release (September 2024) comprised MolmoE-1B, Molmo-7B-O, Molmo-7B-D, and Molmo-72B, trained with Ai2's PixMo data. Molmo 2 (December 2025) added video and multi-image understanding and grounding in 4B and 8B models built on Qwen3 and a 7B model (Molmo2-O-7B) built on Ai2's own Olmo 3. Ai2 has since built specialized variants on the line, including MolmoAct 2 for robot manipulation (May 2026).
Ai2 (Allen Institute for AI) · Molmo family
Molmo2-O-7B is the Molmo 2 vision-language model built on Ai2's Olmo 3 7B Instruct language model with a SigLIP 2 vision encoder, released in December 2025. It accepts images, multiple images, and video with text, and can answer questions, caption, and point to or track objects. Ai2 describes it as the variant whose full model flow is open, since its language model is also Ai2's.
Meta (Meta Superintelligence Labs)
Muse is a model family from Meta Superintelligence Labs, introduced with Muse Spark in April 2026. Muse Spark versions are offered through Meta AI and the Meta Model API without public weights. Muse Glimmer, released in August 2026, is an open-weight model distilled from Muse Spark. Meta's developer site also lists Muse Image and Muse Voice Transcribe.
Microsoft Research
Muse is Microsoft Research's name for its World and Human Action Model (WHAM), a generative model of gameplay trained on human play data from the Xbox game Bleeding Edge. It can generate game visuals, controller actions, or both. It was developed by Microsoft Research's Game Intelligence and Teachable AI Experiences teams with Ninja Theory.
Meta (Meta Superintelligence Labs) · Muse (Meta) family
Muse Glimmer is an open-weight model from Meta Superintelligence Labs with about 29.6 billion parameters, including a roughly 1.8-billion-parameter perception encoder. It is a dense transformer that takes text and images as input and produces text, and was distilled from Muse Spark for agentic tasks on consumer hardware. The model card lists a context length of 131,072+ tokens and a January 4, 2026 knowledge cutoff.
NVIDIA
Nemotron-CC is NVIDIA's English pretraining dataset built from Common Crawl. The original release (December 2024) has 6.3T tokens, 4.4T globally deduplicated original tokens and 1.9T synthetic tokens, drawn from 99 Common Crawl snapshots (CC-MAIN-2013-20 through CC-MAIN-2024-30) and split into five quality buckets. Nemotron-CC-v2 (August 2025) and Nemotron-CC-v2.1 (December 2025), published on Hugging Face, add newer snapshots, further synthetic rephrasings, translated question-answer data, and other subsets.
Nomic AI
Nomic Embed is Nomic's series of open embedding models. It began with nomic-embed-text-v1 (February 2024), a text embedder with an 8,192-token context, and continued with v1.5 and its resizable Matryoshka embeddings (February 2024), vision models aligned to the v1.5 space (June 2024), the multilingual mixture-of-experts nomic-embed-text-v2-moe (February 2025), Nomic Embed Code (March 2025), and Nomic Embed Multimodal (April 2025).
Nomic AI · Nomic Embed family
A text embedding model that updates nomic-embed-text-v1 with Matryoshka representation learning, so its 768-dimension embeddings can be shortened to as few as 64 dimensions, or binarized, with a small loss in quality. Released in February 2024, it follows v1's long-context design, which Nomic describes as supporting 8,192 tokens.
Nomic AI · Nomic Embed family
A multilingual text embedding model that uses a mixture-of-experts design (8 experts, top-2 routing), with 475M total and 305M active parameters. It covers about 100 languages, produces 768-dimensional embeddings that can be truncated to 256 through Matryoshka training, and accepts inputs of up to 512 tokens. Nomic released it in February 2025.
NVIDIA
Cosmos is NVIDIA's family of world foundation models, datasets, and tools for physical AI such as robots, autonomous vehicles, and smart infrastructure. Earlier generations were released as separate Predict, Transfer, and Reason models (for example Cosmos-Predict2.5 and Cosmos-Reason2). Cosmos 3 merges understanding and generation of text, images, video, audio, and actions in one mixture-of-transformers model family; its Super (64B) and Nano (16B) models were released on May 31, 2026, and the Edge (4B) model followed in July 2026.
NVIDIA · NVIDIA Cosmos family
Cosmos3-Nano is the 16B-parameter model in NVIDIA's Cosmos 3 family, released on May 31, 2026. It shares the family's mixture-of-transformers design, generating text autoregressively and images, video, audio, and robot actions by diffusion, from combinations of text, image, video, audio, and action-trajectory inputs.
NVIDIA · NVIDIA Cosmos family
Cosmos3-Super is the 64B-parameter model in NVIDIA's Cosmos 3 family, released on May 31, 2026. It uses a mixture-of-transformers design with an autoregressive tower that generates text and a diffusion tower that generates images, video, audio, and robot actions, and it takes combinations of text, image, video, audio, and action-trajectory inputs.
NVIDIA
Isaac GR00T N is NVIDIA's family of open vision-language-action models for humanoid and other robots. The models take camera images, language instructions, and robot state and output continuous actions, using a vision-language backbone with a diffusion-transformer action head. The repository describes N1.7 as the latest version of GR00T N1 and lists N1.5 and N1.6 as previous releases.
NVIDIA · NVIDIA Isaac GR00T N family
GR00T N1.7 is a 3B-parameter vision-language-action model from NVIDIA that maps camera images, language instructions, and robot proprioception to continuous robot actions. It replaces the earlier Eagle backbone with Cosmos-Reason2-2B and uses a flow-matching diffusion transformer as its action head. NVIDIA first released it in early access in April 2026; the repository now describes it as a general availability release.
NVIDIA
NeMo Framework is NVIDIA's open-source Python framework for pretraining, post-training, and reinforcement learning of generative AI models, organized as separate libraries in the NVIDIA-NeMo GitHub organization, including Megatron Bridge, AutoModel, NeMo RL, Curator, Evaluator, Export-Deploy, Run, and NeMo Speech. In 2026 the original NeMo repository was split and now focuses on audio, speech, and multimodal LLMs as NeMo Speech (version 3.0), and for the framework container from 26.02 onward NVIDIA points to Megatron Bridge's documentation and release notes.
NVIDIA
Nemotron is NVIDIA's family of models for building AI agents. The current generation is organized into Nano, Lightning, Super, and Ultra tiers; NVIDIA publishes the weights on Hugging Face and, for several releases, training datasets and training recipes.
NVIDIA · NVIDIA Nemotron family
Nemotron 3 Super is an NVIDIA language model with 120B total and 12B active parameters, using a hybrid Mamba-2 and attention architecture with latent mixture-of-experts layers and multi-token prediction, and configurable reasoning. NVIDIA released BF16, FP8, and NVFP4 checkpoints and a separate base checkpoint on Hugging Face.
NVIDIA · NVIDIA Nemotron family
Nemotron 3 Ultra is an NVIDIA language model with 550B total and 55B active parameters, built on a hybrid Mamba-2 and attention architecture with latent mixture-of-experts layers and multi-token prediction. Reasoning can be switched on or off through the chat template. NVIDIA released BF16 and NVFP4 checkpoints and a separate base checkpoint on Hugging Face.
NVIDIA · NVIDIA Nemotron family
Nemotron 3.5 Lightning is an NVIDIA language model with 30B total and 3B active parameters, using a hybrid Mamba-2 and attention mixture-of-experts architecture with multi-token prediction and configurable reasoning. NVIDIA released full-precision BF16 weights, an NVFP4 checkpoint for deployment, draft models for speculative decoding, and a separate base checkpoint.
Ollama Inc.
Ollama is an open-source application and runtime for downloading and running open models on macOS, Windows, and Linux. It has an API, Python and JavaScript client libraries, and integrations with coding agents and other tools. Its company, Ollama Inc., also offers cloud-hosted models.
Ai2 (Allen Institute for AI)
OLMo (styled Olmo in recent releases) is Ai2's series of open language models. Releases have included the first OLMo 1B and 7B models (February 2024), OLMoE (September 2024), OLMo 2 (November 2024 to May 2025), and Olmo 3 at 7B and 32B (November 2025), with Olmo 3.1 32B checkpoints added in December 2025. In March 2026 Ai2 also released Olmo Hybrid, a 7B model that mixes attention layers with linear RNN layers.
Ai2 (Allen Institute for AI) · OLMo family
The pretrained base model at the 7B scale in Ai2's Olmo 3 release, trained in three stages (pretraining, mid-training, and long-context extension) on Dolma 3 data. The model card lists 5.93 trillion training tokens, 32 layers, and a 65,536-token context length.
Ai2 (Allen Institute for AI) · OLMo family
A 7B chat model in Ai2's Olmo 3 release, post-trained from the Olmo 3 7B base model with supervised fine-tuning, direct preference optimization, and reinforcement learning with verifiable rewards on the Dolci datasets. It responds without a separate reasoning trace, unlike the Olmo 3 Think models.
Ai2 (Allen Institute for AI) · OLMo family
A 32B reasoning model from Ai2's December 2025 Olmo 3.1 update. It was produced by continuing the reinforcement learning run behind Olmo 3 32B Think for about three more weeks with additional passes over the Dolci-Think-RL data, and it writes a reasoning trace before its final answer.
ONNX project (hosted by the LF AI & Data Foundation)
ONNX is an open format for representing machine learning models, covering both deep learning and traditional ML. It defines an extensible computation graph model, built-in operators, and standard data types, with a current focus on inference. Facebook and Microsoft started the ONNX community in 2017, and LF AI announced it as a graduate-level project in November 2019.
Microsoft
ONNX Runtime is Microsoft's open-source, cross-platform engine for running and accelerating machine-learning models in the ONNX format. It runs models exported from frameworks such as PyTorch and TensorFlow/Keras and from classical libraries such as scikit-learn, LightGBM, and XGBoost, applying graph optimizations and hardware-specific execution providers; it also offers accelerated training for transformer models on NVIDIA GPUs.
OpenAI
OpenAI Evals is an open-source framework from OpenAI for evaluating language models and systems built on them, together with a registry of existing evals. Evals are defined in YAML and run from the oaieval and oaievalset command-line tools against "completion functions", which by default call models through the OpenAI API. The repository is not archived, but activity since mid-2024 has been limited to maintenance, and its README now points users to OpenAI's hosted Evals in the OpenAI dashboard.
Apple
OpenELM is a family of small language models released by Apple in April 2024 in 270M, 450M, 1.1B, and 3B parameter sizes, each as a pretrained and an instruction-tuned model. The models use layer-wise scaling to allocate parameters across transformer layers, and Apple published them together with its CoreNet training library, pre-training configurations, training logs, and evaluation instructions.
Apple · OpenELM family
OpenELM 3B Instruct is the 3B-parameter instruction-tuned model in Apple's April 2024 OpenELM release. It was pretrained with Apple's CoreNet library on about 1.8 trillion tokens of public data and then instruction-tuned on the UltraFeedback dataset using the Hugging Face Alignment Handbook.
All Hands AI (OpenHands)
OpenHands is an open-source platform for building and running AI coding agents. Its main repository now holds Agent Canvas, a self-hosted browser control center. Agent Canvas runs the OpenHands agent, or other agents that support the Agent Client Protocol, on local, remote, or cloud backends. The agent itself, the Python SDK, and the Agent Server are developed in the separate Software Agent SDK repository.
Physical Intelligence
openpi is Physical Intelligence's repository of code and model checkpoints for its π0, π0-FAST, and π0.5 vision-language-action models for robot control. It provides base checkpoints pre-trained on more than 10,000 hours of robot data, fine-tuned checkpoints for the ALOHA, DROID, and LIBERO setups, and JAX and PyTorch code for inference and fine-tuning.
OpenXLA project maintainers (Google-led technical leadership)
OpenXLA is an open-source project that develops machine learning compiler and infrastructure components linking frameworks such as JAX, PyTorch, and TensorFlow to hardware backends. Its website lists XLA (a compiler for GPUs, CPUs, and ML accelerators), StableHLO (a portable operation set), Shardy (an MLIR-based tensor partitioner), PJRT (a device and compiler interface), XProf (profiling tools), and Tokamax.
Writer
Palmyra is Writer's family of language models. The Palmyra X models (currently X4, X5, and X6) are offered through Writer's platform and API. Writer has also published Palmyra models on Hugging Face: Palmyra-X-4.3-73B and domain models for medicine and finance under its non-commercial open model license, and the small Palmyra-mini models (September 2025) under Apache 2.0.
Writer · Palmyra family
A small text model from Writer, fine-tuned from Qwen2.5-1.5B, that the model card lists at 1.7 billion parameters with a 131,072-token context window. It is the non-reasoning member of the three-model Palmyra-mini family that Writer released in September 2025.
Writer · Palmyra family
A small reasoning model in Writer's Palmyra-mini family, released in September 2025. Writer built it on NVIDIA's OpenReasoning-Nemotron-1.5B and applied reinforcement-learning fine-tuning; the model card lists a 131,072-token context window.
Microsoft
Phi is Microsoft's family of small language models. Generations have included Phi-3, Phi-3.5, and Phi-4, with instruction, reasoning, multimodal, and vision-language variants published on Hugging Face and Microsoft Foundry. Recent Phi-4 releases include Phi-4-mini-flash-reasoning (July 2025) and Phi-4-reasoning-vision-15B (March 2026).
Microsoft · Phi family
Phi-4-mini-flash-reasoning is a 3.8-billion-parameter, English, text-only Microsoft model fine-tuned for mathematical reasoning. It uses a hybrid "SambaY" decoder-hybrid-decoder architecture that mixes state space model layers with attention and Differential Attention, and supports a 64K-token context.
Microsoft · Phi family
Phi-4-reasoning-vision-15B is a 15-billion-parameter multimodal reasoning model from Microsoft. It combines the Phi-4-Reasoning language model with a SigLIP-2 vision encoder, takes text and images as input, produces text, and has a 16,384-token context length. It can either reason step by step or answer directly, depending on the task.
Prime Intellect
prime-rl is Prime Intellect's open-source framework for large-scale, asynchronous reinforcement learning and supervised fine-tuning of language models. It separates an FSDP2-based trainer, a vLLM inference service, and an orchestrator that gathers rollouts from verifiers environments, and it supports multi-node deployment on Slurm and Kubernetes.
EleutherAI
Pythia is EleutherAI's suite of language models built for research on how models learn during training and change with scale. The main suite has eight sizes from 70M to 12B parameters, each trained once on the Pile and once on a deduplicated Pile, on the same data in the same order, with 154 checkpoints per model. The current suite was retrained and re-released in April 2023; 14M and 31M models and additional random-seed runs were added later.
EleutherAI · Pythia family
The largest model in EleutherAI's Pythia suite: an English decoder-only language model with about 11.8 billion total parameters, trained on the Pile (not deduplicated) for about 300 billion tokens. EleutherAI publishes 154 checkpoints from initialization to the end of training as Hugging Face branches.
PyTorch Foundation
PyTorch is an open-source machine learning framework for tensors and dynamic neural networks in Python, with GPU acceleration and a C++ distribution (LibTorch). Meta originally developed it and released it in 2016. Since September 2022 it has been hosted by the PyTorch Foundation under the Linux Foundation.
Lightning AI
PyTorch Lightning is an open-source Python framework that structures PyTorch training code and handles engineering tasks such as multi-GPU and multi-node distributed training and mixed precision. The same repository contains Lightning Fabric, a lighter layer for scaling custom PyTorch training loops, and both ship in the lightning package. It was open-sourced in 2019 and is maintained by Lightning AI.
Qualcomm Technologies, Inc.
A Python package and GitHub repository of machine learning models prepared for deployment on Qualcomm devices. It provides export scripts that compile, quantize where applicable, profile, and run models through Qualcomm AI Hub Workbench, end-to-end demos for most models, and sample application code. Many of the packaged models come from other developers, such as YOLO, Llama, Mistral, Phi, and Qwen variants.
PyTorch Foundation (hosted project)
Ray is an open-source distributed computing framework for AI workloads, with a core distributed runtime and libraries for data processing, training, and serving. It was developed at UC Berkeley's RISELab and became a PyTorch Foundation-hosted project in October 2025, contributed by Anyscale.
Together AI (Together Computer, Inc.)
RedPajama is Together AI's open pretraining dataset project. RedPajama-V1 (RedPajama-Data-1T, 2023) is a 1.2-trillion-token reproduction of the LLaMA training-data recipe drawn from Common Crawl, C4, GitHub, arXiv, Wikipedia, and Stack Exchange. The current version, RedPajama-V2 (October 2023), contains over 100 billion documents from 84 Common Crawl snapshots in English, German, French, Spanish, and Italian, with more than 40 precomputed quality signals and duplicate markers; its deduplicated, annotated portion is about 30 trillion tokens.
Essential AI
Rnj is Essential AI's line of 8B-parameter dense language models trained from scratch and aimed at code and STEM work. Rnj-1 (a base model and an instruction-tuned model) was released in December 2025, followed in 2026 by Rnj-1.5 Instruct, which extends the context window from 32K to 160K tokens.
Essential AI · Rnj family
The instruction-tuned version of Essential AI's Rnj-1, an 8.3B-parameter dense model trained from scratch. The base model was pretrained on 8.4T tokens at an 8K context, extended to a 32K context in a 380B-token mid-training stage, and then given a 150B-token supervised fine-tuning stage to produce this model.
Essential AI · Rnj family
A long-context follow-up to Rnj-1 Instruct that extends the context window from 32K to 160K tokens. Essential AI built it from the Rnj-1 base model, switching most attention layers to block-local attention with a group of global layers in the middle, and added long-context mid-training data and more software-engineering training data.
Meta (FAIR) · Segment Anything (SAM) family
SAM 2.1 Hiera-Large is the largest checkpoint (224.4M parameters, per the README) in Meta's SAM 2.1 suite, an improved set of SAM 2 checkpoints released in September 2024. SAM 2 is a transformer with streaming memory that segments objects in images and tracks them through video from point, box, or mask prompts.
Meta (Meta Superintelligence Labs) · Segment Anything (SAM) family
SAM 3.1 is a March 2026 update of Meta's SAM 3, a model that detects, segments, and tracks objects in images and video from text phrases, image exemplars, or visual prompts such as points, boxes, and masks. The update adds Object Multiplex, which groups tracked objects into shared-memory buckets and processes them jointly instead of one at a time, together with new checkpoints and inference optimizations.
Meta
Segment Anything is Meta's family of promptable segmentation models. The original SAM (April 2023) produces object masks in images from prompts such as points or boxes; SAM 2 (July 2024, updated as SAM 2.1 in September 2024) extends this to video with streaming memory; SAM 3 (November 2025, updated as SAM 3.1 in March 2026) adds detection, segmentation, and tracking of every instance of a concept named by a short text phrase or image exemplar. Licenses differ by generation: Apache 2.0 for SAM and SAM 2, and Meta's custom SAM License for SAM 3.
LMSYS (hosting organization), RadixArk (company described as a maintainer in a post on the LMSYS blog)
SGLang is an open-source serving framework for large language models and multimodal models, designed for low-latency, high-throughput inference on anything from a single GPU to large distributed clusters. Its runtime features include RadixAttention prefix caching, prefill-decode disaggregation, speculative decoding, continuous batching, several forms of parallelism, structured outputs, and quantization.
Hugging Face (Smol Models Research, HuggingFaceTB)
SmolLM is Hugging Face's series of small language models. The first SmolLM release (July 2024) came in 135M, 360M, and 1.7B sizes; SmolLM2 kept those three sizes, with the 1.7B model trained on about 11 trillion tokens (paper February 2025); and SmolLM3 (July 2025) is a 3B model released as base and instruct checkpoints, with a dual-mode reasoning instruct model. The SmolVLM vision-language models are maintained in the same repository.
Hugging Face (Smol Models Research, HuggingFaceTB) · SmolLM family
SmolLM3-3B is the instruct model of Hugging Face's SmolLM3 release (July 2025), a 3B-parameter decoder-only transformer pretrained on about 11 trillion tokens of web, code, math, and reasoning data in stages, then mid-trained on reasoning data and aligned with supervised fine-tuning and Anchored Preference Optimization. It can answer with or without an extended reasoning trace, supports tool calling, and natively covers six European languages.
Snowflake AI Research
Arctic is a text-in, text-and-code-out language model pretrained from scratch by the Snowflake AI Research Team. It uses a dense-MoE hybrid transformer that combines a 10B dense model with a residual 128x3.66B mixture-of-experts MLP, for 480B total and 17B active parameters with top-2 gating. Base and instruct-tuned checkpoints were released on April 24, 2024 with weights and code under Apache-2.0.
Snowflake AI Research · Snowflake Arctic family
Arctic Base is the pretrained checkpoint of Snowflake Arctic, a dense-MoE hybrid transformer (a 10B dense model plus a residual 128x3.66B MoE MLP) with 480B total and 17B active parameters, pretrained from scratch by the Snowflake AI Research Team. It takes text input and generates text and code.
Snowflake AI Research · Snowflake Arctic family
Arctic Instruct is the instruct-tuned checkpoint of Snowflake Arctic, a dense-MoE hybrid transformer (a 10B dense model plus a residual 128x3.66B MoE MLP) with 480B total and 17B active parameters, developed by the Snowflake AI Research Team. It takes text input and generates text and code.
BigCode project (open scientific collaboration), Hugging Face (BigCode co-steward), ServiceNow (BigCode co-steward)
StarCoder2 is a family of code generation models from the BigCode project, released on February 28, 2024 in 3B, 7B, and 15B sizes and trained on The Stack v2. Per Hugging Face's announcement, ServiceNow trained the 3B model, Hugging Face the 7B model, and NVIDIA the 15B model. The bigcode organization on Hugging Face also hosts an instruction-tuned variant, StarCoder2-15B-Instruct-v0.1.
BigCode project (open scientific collaboration), Hugging Face (BigCode co-steward), ServiceNow (BigCode co-steward) · StarCoder2 family
The largest StarCoder2 base model, with 15B parameters, trained on more than 4 trillion tokens covering 600+ programming languages from The Stack v2. It uses grouped-query attention, a 16,384-token context window with 4,096-token sliding-window attention, and a fill-in-the-middle training objective. NVIDIA trained it with the NeMo framework on its Eos supercomputer.
BigCode project (open scientific collaboration), ServiceNow (BigCode co-steward; trained this model), Hugging Face (BigCode co-steward) · StarCoder2 family
The smallest StarCoder2 base model, with about 3B parameters, trained on more than 3 trillion tokens from The Stack v2 and other sources. Its model card lists 17 programming languages, grouped-query attention, a 16,384-token context window with 4,096-token sliding-window attention, and a fill-in-the-middle objective. ServiceNow trained it for the BigCode project.
ServiceNow (SLAM Labs) · Apriel family
A 15B-parameter instruction-tuned "supernet" in which each of 48 decoder layers carries four token-mixer variants: full attention, sliding-window attention, Gated DeltaNet, and Kimi Delta Attention. Choosing one mixer per layer gives deployment presets that trade output quality for decoding speed from a single checkpoint. It was derived from Apriel-1.6-15b-Thinker by distillation followed by supervised fine-tuning.
Google (TensorFlow team)
TensorFlow is an open-source, end-to-end machine learning platform with stable Python and C++ APIs, originally developed within Google Brain. With the 2.21 release in March 2026, Google said it would limit TensorFlow work to security and bug fixes, dependency updates, and community contributions, and pointed new generative AI work to Keras 3, JAX, and PyTorch.
NVIDIA
TensorRT-LLM is NVIDIA's open-source library for running large language model inference on NVIDIA GPUs. It provides a Python LLM API, optimized kernels, Python and C++ runtimes, and an online serving command (trtllm-serve).
Stanford University, Laude Institute, Harbor (harbor-framework project)
Terminal-Bench is a benchmark of hard tasks that AI agents must complete in command-line environments; each task has its own container environment, a human-written reference solution, and verification tests. It was first released in May 2025; Terminal-Bench 2.0 (November 2025, 89 tasks) introduced the Harbor framework for running tasks, and since July 2026 it has been maintained as a "continuous benchmark" with versioned releases, the latest being 4.0 (August 2026).
EleutherAI
The Pile is an 825 GiB English text dataset for language-model pretraining, assembled by EleutherAI from 22 component datasets that include web text (Pile-CC, from Common Crawl), academic and professional sources such as PubMed Central, arXiv, and FreeLaw, books, GitHub code, and dialogue. It was described in a December 2020 paper and a 2022 datasheet and is the training corpus of EleutherAI's Pythia models. Its current availability is limited; see access conditions.
Google Research
TimesFM (Time Series Foundation Model) is a family of pretrained time-series forecasting models developed by Google Research, first described in an ICML 2024 paper. Open checkpoints have been released as versions 1.0, 2.0, 2.5 (September 2025), and 3.0 (August 2026); version 3.0 adds native multivariate forecasting and native support for past-only and past-and-future covariates.
Google Research · TimesFM family
TimesFM 2.5 is a 200M-parameter time-series forecasting model from Google Research, released in September 2025. Compared with TimesFM 2.0 it has fewer parameters (200M, down from 500M), a context length of up to 16k points, an optional 30M quantile head for continuous quantile forecasts up to a 1k horizon, and no frequency indicator input.
Google Research · TimesFM family
TimesFM 3.0 is a Google Research time-series forecasting model released in August 2026. Google's announcement gives it 330 million parameters and describes native multivariate forecasting, support for past-only and past-and-future covariates, and probabilistic forecasts with nine quantiles. The model card lists a 20-layer transformer architecture.
Hugging Face
Transformers is Hugging Face's open-source Python library of model definitions for text, vision, audio, video, and multimodal models, used for both inference and training. Its README presents it as a shared model-definition layer that other training frameworks and inference engines build on, and it loads pretrained checkpoints from the Hugging Face Hub.
Arcee AI
Trinity is Arcee AI's family of sparse mixture-of-experts language models: Trinity Nano (6B total, about 1B active parameters), Trinity Mini (26B total, 3B active), and Trinity Large (about 398B total, 13B active). Trinity Large is published as base, pre-anneal base, preview, and reasoning-tuned Thinking checkpoints.
Arcee AI · Trinity family
Trinity Mini is a 26B-parameter mixture-of-experts model with 3B active parameters, 128 experts (8 active plus 1 shared), and a 128k-token context window. It was trained on 10 trillion tokens and tuned for reasoning.
Arcee AI · Trinity family
Trinity-Large-Thinking is a reasoning-tuned sparse mixture-of-experts model with about 398B total and 13B active parameters per token and a 512k-token context window. It was post-trained from Trinity-Large-Base with chain-of-thought and agentic reinforcement learning, and writes its reasoning in think blocks before answering.
Hugging Face
TRL is Hugging Face's open-source library for post-training transformer language models. It provides trainer classes for methods including supervised fine-tuning (SFT), Group Relative Policy Optimization (GRPO), Direct Preference Optimization (DPO), KTO, and reward modeling, with further methods marked experimental, and it is built on the Transformers library.
Ai2 (Allen Institute for AI)
Tülu is Ai2's line of instruction-following models together with the open post-training recipe used to produce them. Tülu 3 (November 2024) applied supervised fine-tuning, Direct Preference Optimization, and reinforcement learning with verifiable rewards (RLVR) to Meta's Llama 3.1 base models at 8B and 70B, followed by a 405B model in January 2025 and by Tülu 3.1 8B, which changed only the final RL stage (from PPO to GRPO).
Meta (FAIR)
V-JEPA 2 is Meta's family of self-supervised video encoders, pretrained on video with a masked latent-feature prediction objective. Meta released it in June 2025 with ViT-L, ViT-H, and ViT-g encoders and an action-conditioned world model (V-JEPA 2-AC) post-trained on robot video, and in March 2026 added V-JEPA 2.1, a retrained family (ViT-B to a 2B-parameter ViT-G) aimed at dense, temporally consistent features.
Meta (FAIR) · V-JEPA 2 family
The 1-billion-parameter ViT-g/16 encoder from Meta's V-JEPA 2, in the version that takes 64-frame clips at 384-pixel resolution. It was pretrained with self-supervision on VideoMix22M, a mix of about 22 million video and image samples. Meta's action-conditioned V-JEPA 2-AC world model was post-trained from a ViT-g/16 V-JEPA 2 encoder.
Meta (FAIR) · V-JEPA 2 family
The largest V-JEPA 2.1 model, a 2-billion-parameter ViT-G/16 video and image encoder released by Meta in March 2026. V-JEPA 2.1 changes the V-JEPA 2 recipe to learn dense, temporally consistent features, using a dense predictive loss over all tokens, self-supervision at several intermediate layers, and separate tokenizers for images and videos. Meta distilled this model into smaller ViT-B and ViT-L variants.
PyTorch Foundation (hosted project), vLLM core maintainers (project Technical Steering Committee)
vLLM is an open-source inference and serving engine for large language models, originally built around the PagedAttention memory-management technique. It was developed in UC Berkeley's Sky Computing Lab, and UC Berkeley contributed it to the PyTorch Foundation in 2025.
Waymo
Waymax is a multi-agent driving simulator written in JAX that builds scenarios from the Waymo Open Motion Dataset and represents road users as bounding boxes rather than raw sensor data. It includes data loading, visualization, metrics, simulated agents, and adapters to reinforcement learning interfaces such as dm-env and Brax.
Waymo
Autonomous driving data collected by Waymo vehicles, released for research in August 2019 and since expanded into three datasets: Perception (lidar and camera data with labels for 2,030 segments), Motion (object trajectories and 3D maps for 103,354 segments), and End-to-End Driving (360-degree camera images and routing instructions for 5,000 segments). A companion repository provides the data format definitions, evaluation metrics, and TensorFlow helper code.
Microsoft Research · Muse (Microsoft Research WHAM) family
WHAM 1.6B is the larger of the two World and Human Action Model checkpoints Microsoft Research released for Muse. It pairs a VQ-GAN image tokenizer with a transformer trained from scratch to predict game visuals and controller actions. The same repository includes a 200M-parameter checkpoint for lighter experiments.
OpenAI
Whisper is OpenAI's family of encoder-decoder transformer models for speech recognition, speech translation into English, and language identification. The original series of English-only and multilingual checkpoints from 39M to 1550M parameters appeared in September 2022; OpenAI followed with large-v2 (December 2022), large-v3 (November 2023), and the faster large-v3-turbo (September 2024).
OpenAI · Whisper family
Whisper large-v3 is a 1550M-parameter multilingual speech recognition and translation model released by OpenAI in November 2023. It keeps the architecture of the earlier large models but uses 128 Mel frequency bins instead of 80 and adds a language token for Cantonese.
OpenAI · Whisper family
Whisper large-v3-turbo is a multilingual speech recognition model that OpenAI derived from Whisper large-v3 by cutting the decoder from 32 layers to 4 and fine-tuning for two more epochs. It has about 0.8 billion parameters (809M in the README and Hugging Face card; the repository model card lists 798M) and is the default model in the openai-whisper package.
Zyphra
Zamba is Zyphra's series of hybrid models that combine Mamba state-space layers with shared transformer attention blocks. Releases include Zamba-7B (April 2024); Zamba2 at 1.2B, 2.7B, and 7B parameters (July to October 2024), with instruction-tuned versions; and Zamba2-VL vision-language models at the same three sizes (June 2026).
Zyphra · Zamba family
The 7B base model in Zyphra's Zamba2 series, announced in October 2024. It interleaves Mamba2 blocks with two shared attention blocks in an alternating pattern and adds LoRA projectors to the shared MLP blocks so that each shared block can specialize by depth.
Zyphra · Zamba family
The largest of the three Zamba2-VL vision-language models Zyphra released in June 2026, with about 8B parameters. It pairs the Zamba2-7B hybrid state-space language model with the Qwen2.5-VL vision encoder and handles single- and multi-image understanding and grounding.
Sierra
τ-bench is Sierra's open-source simulation framework for evaluating customer-service AI agents. In each domain an agent must follow a written policy and use tools while a simulated user takes part in the conversation, in turn-based text mode or full-duplex voice mode. The current repository, which carries the τ³-bench update, covers airline, retail, telecom, and banking-knowledge domains plus a mock domain.
Help keep USASI useful.
Optional. No USASI account required. Payment takes place on the linked provider’s website (Buy Me a Coffee).
About supporting this project