LFM2.5-8B-A1B
Release in the LFM (Liquid Foundation Models) family · version LFM2.5-8B-A1B
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.14
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- License: License (LFM Open License v1.0) (external site: huggingface.co)
- Release notes: Release post (external site: liquid.ai)
Availability and license
Overall availability
Weights download from Hugging Face without an access gate; use is governed by the LFM Open License v1.0, which conditions commercial-use rights on the licensee's legal entity not exceeding a threshold defined as annual revenue of US$10,000,000 or more. Liquid also offers the model in its Playground.124
Availability is separate from permission: read the license before using or redistributing.
Custom Liquid AI license (card license field "lfm1.0") modeled on Apache 2.0. It grants perpetual, royalty-free copyright and patent licenses, requires redistributors to include the license and mark modified files, and ends patent rights for anyone who files patent litigation over the work. Section 5 makes commercial rights conditional on the licensee's legal entity not exceeding a "Threshold" defined as annual revenue of US$10,000,000 or more, and states that commercial use by a legal entity that exceeds the Threshold is not licensed. The Threshold does not apply to a qualified non-profit organization's use for non-commercial or research purposes. Liquid's license page FAQ says fine-tuned models may be kept proprietary and that hosting platforms may host and distribute the models. The release post describes the weights as deployable without restrictions; the license text limits commercial use as described here.234
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Ungated safetensors weights on Hugging Face; GGUF, ONNX, and MLX conversions are published as separate repositories.1 |
| Inference codeIs code for running the model published? | Public | The model card documents inference with Hugging Face Transformers (5.0.0 or later), vLLM, SGLang, llama.cpp, MLX, and LM Studio; no custom modeling code is needed in the repository.1 |
| Training codeIs the code used to train the model published? | Unknown | No training code release was found in the model card, release post, or LFM2 technical report. Liquid documents fine-tuning with third-party tools (Unsloth, TRL), which is not training code for this release. |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Unknown | The model card and release post give the 38-trillion-token pretraining budget but do not describe data sources or composition. |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The release post outlines the stages at a high level (scaled-up pretraining, context extension to 32K and then 128K tokens, reasoning training, preference optimization against repetitive loops, and reward-based hallucination mitigation) without full configurations.4 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The model card and release post report benchmark tables; evaluation code or prompts are not linked.14 |
What it is useful for
The model card recommends it for agentic workflows, tool use, structured outputs, multilingual assistants, and on-device personal-assistant applications, and says it is not the best fit for heavy programming or knowledge-intensive question answering without retrieval.1
Run and use notes
- The model card recommends temperature 0.2, top_k 80, and repetition_penalty 1.05, and shows Transformers loading in bfloat16. It also points to a separate 328M speculative decoding drafter, LFM2.5-8B-A1B-DSpark, for use with SGLang.1
Organization context
Provenance and derivatives
Post-trained by Liquid AI from its own LFM2.5-8B-A1B-Base checkpoint, which it pretrained itself; no third-party base model is involved.1
- Derived from: LFM2.5-8B-A1B-Base (external site: huggingface.co) — Same organization; pretrained base checkpoint.
Other releases in the LFM (Liquid Foundation Models) family
- LFM2.5-2.6BModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.