INTELLECT-3
Release in the INTELLECT family · version INTELLECT-3
Maintained by Prime Intellect13
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.134
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Paper: INTELLECT-3 technical report (arXiv 2512.16144) (external site: arxiv.org)
- Release notes: Release post (external site: primeintellect.ai)
- Repository: prime-rl (training framework) (external site: github.com)
Availability and license
Overall availability
Weights download from Hugging Face without an access gate; an FP8 version is published separately. The release post also points to a chat interface and to hosting by third-party providers.14
Availability is separate from permission: read the license before using or redistributing.
Apache License 2.0 (prime-rl) (external site: raw.githubusercontent.com)5
MIT License (verifiers) (external site: raw.githubusercontent.com)6
The MIT license for the weights is declared in the model card metadata; the repository has no separate LICENSE file. The card says the model, training frameworks, and environments are released under MIT and Apache 2.0 licenses. The base model, GLM-4.5-Air-Base, is also tagged MIT on its model card.127
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 | BF16 safetensors in an ungated Hugging Face repository; an FP8 repository is also published.1 |
| Inference codeIs code for running the model published? | Public | The model card gives vLLM serving commands with the qwen3_coder tool-call parser and deepseek_r1 reasoning parser.1 |
| Training codeIs the code used to train the model published? | Partial | The SFT and RL stages were run with prime-rl (Apache 2.0) using environments built with the verifiers library (MIT), both open source. This covers Prime Intellect's post-training only; the code used to pretrain the GLM-4.5-Air base is not part of this release.356 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The report lists the SFT sources with sample and token counts (splits of NVIDIA's Nemotron-Post-Training-Dataset-v1, AM-DeepSeek-R1-0528-Distilled, SWE-Swiss, Toucan Tool, and synthetic data from Environments Hub environments) and the datasets behind each RL environment. The pretraining data of the base model is not described in this release.3 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | For the post-training, the report gives SFT settings (Muon optimizer, learning rates, context lengths of 65K and 98K, FSDP and context parallelism) and RL settings (256 prompts with 16 rollouts each, 65,536-token context, learning rate 1e-6, 60 nodes split between training and inference, masked importance sampling). The base model's pretraining recipe is Z.ai's and is not part of this release.3 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The model card and report give benchmark results, and the report's appendix documents the evaluation setup and names the Environments Hub environments used. The hub pages opened only as a sign-in dashboard during review, so public access to the evaluation code was not confirmed.13 |
What it is useful for
Run and use notes
- The model card states that the BF16 version can be served with vLLM on 2x H200 GPUs (tensor parallel size 2) and the FP8 version on a single H200.1
Organization context
Provenance and derivatives
Post-trained by Prime Intellect from GLM-4.5-Air-Base, a 106B-total, 12B-active base model released by Z.ai (zai-org) under the MIT License. The base model is not a catalog member and is not treated as U.S.-developed; Prime Intellect's contribution is the SFT and RL training, the environments, and the training framework.137
- Derived from: GLM-4.5-Air-Base (Z.ai) (external site: huggingface.co) — Third-party base model; not a catalog member.
Catalog records that name this entry in their provenance:
Other releases in the INTELLECT family
- INTELLECT-3.1Model-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: U.S. headquarters
Post-trained and published by Prime Intellect, Inc., a Delaware corporation with U.S. locations (see the prime-intellect record). Eligibility covers Prime Intellect's post-training only; the GLM-4.5-Air-Base model it starts from was released by Z.ai and is not treated as a U.S.-developed model.187
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.