Apriel-1.6-15B-Thinker
Release in the Apriel family · version Apriel-1.6-15b-Thinker
Maintained by ServiceNow (SLAM lab)13
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.132
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Release notes: Release blog post (external site: huggingface.co)
- Paper: Apriel-1.5-15b-Thinker technical report (arXiv 2510.01141) (external site: arxiv.org)
- Repository: Fast-LLM training library (external site: github.com)
Availability and license
Overall availability
Downloadable from Hugging Face without gating.12
Availability is separate from permission: read the license before using or redistributing.
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.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Safetensors weights are downloadable from Hugging Face; a GGUF conversion is published in a separate repository.126 |
| Inference codeIs code for running the model published? | Public | The model card gives inference code for Hugging Face Transformers and a vLLM server command with Apriel tool-call and reasoning parsers.1 |
| Training codeIs the code used to train the model published? | Partial | The card names Fast-LLM and VERL as the training stack. Fast-LLM is ServiceNow's public training library (Apache 2.0), but this catalog did not find the Apriel-1.6 training configurations published.17 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The card and blog describe data categories (math, code, science, logical reasoning, multimodal image-text data, 2.4M SFT samples, synthetic data, and a share of NVIDIA Nemotron datasets in the upscaling mix). The training datasets themselves are not documented as released.13 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The card and blog outline the stages (depth-upscaling and continual pretraining, text then multimodal SFT, and multi-stage RL with GSPO) at a high level.13 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The card publishes a results table; some text benchmark scores are taken from Artificial Analysis and the rest were evaluated internally. No evaluation code for this release was found.1 |
What it is useful for
The card lists code assistance and generation, logical and multi-step reasoning, question answering and information retrieval, function calling, complex instruction following, and agent use cases. It says the model is not intended for safety-critical applications without human oversight.1
Run and use notes
- The model card documents inference with Hugging Face Transformers in bfloat16 and a vLLM OpenAI-compatible server with a 131,072-token maximum model length, and recommends a temperature of 0.6. It says the vLLM parsers were not yet merged upstream at the time and points to a custom Docker image.1
- The card lists Together AI for hosted inference and Ollama for local use as additional options.1
Organization context
Provenance and derivatives
Apriel-1.6-15b-Thinker builds on ServiceNow's Apriel-1.5-15b-Thinker. The Apriel 1.5 report says that model started from Mistral AI's Pixtral-12B-Base-2409 (using a copy hosted by Unsloth), which ServiceNow depth-upscaled from 40 to 48 decoder layers before continued pretraining. The base weights were therefore not developed by ServiceNow.1345
- Derived from: Apriel-1.5-15b-Thinker (external site: huggingface.co) — Direct predecessor; no separate catalog record.
- Derived from: Pixtral-12B-Base-2409 (Mistral AI) (external site: huggingface.co) — Starting checkpoint of the Apriel 1.5 line; Apache 2.0.
Catalog records that name this entry in their provenance:
Other releases in the Apriel family
- SuperApriel-15B-InstructModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: U.S.-governed project
The model is published by ServiceNow's ServiceNow-AI organization on Hugging Face and described as part of ServiceNow's Apriel series. ServiceNow, Inc. is a Delaware corporation headquartered in Santa Clara, California, per its 2025 Form 10-K. The Apriel 1.5 line it builds on started from Pixtral-12B-Base-2409, published by Mistral AI; that base model is recorded under provenance and is not treated as U.S.-developed.184
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.