README.md
3.5 KB · 100 lines · markdown Raw
1 ---
2 library_name: transformers
3 tags:
4 - prime-rl
5 - verifiers
6 - prime-intellect
7 - reinforcement-learning
8 - reasoning
9 - agentic
10 - mixture-of-experts
11 license: mit
12 language:
13 - en
14 base_model:
15 - zai-org/GLM-4.5-Air-Base
16 pipeline_tag: text-generation
17 ---
18
19 # INTELLECT-3
20
21 <div align="center">
22 <img src="banner.png" alt="Prime Intellect Logo" />
23 </div>
24
25 <p align="center">
26 <strong>INTELLECT-3: A 100B+ MoE trained with large-scale RL</strong>
27 <br><br>
28 Trained with <a href="https://github.com/PrimeIntellect-ai/prime-rl">prime-rl</a> and <a href="https://github.com/PrimeIntellect-ai/verifiers">verifiers</a>
29 <br>
30 Environments released on <a href="https://app.primeintellect.ai/dashboard/environments">Environments Hub</a>
31 <br>
32 Read the <a href="https://primeintellect.ai/blog/intellect-3">Blog</a> & <a href="https://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf">Technical Report</a>
33 <br>
34 <a href="https://x.com/primeintellect">X</a> | <a href="https://discord.gg/RC5GvMbfDf">Discord</a> | <a href="https://app.primeintellect.ai/dashboard/create-cluster">Prime Intellect Platform</a>
35 </p>
36
37 ## Introduction
38
39 **INTELLECT-3** is a 106B (A12B) parameter Mixture-of-Experts reasoning model post-trained from [GLM-4.5-Air-Base](https://huggingface.co/zai-org/GLM-4.5-Air-Base) using supervised fine-tuning (SFT) followed by large-scale reinforcement learning (RL).
40
41 ![bench](bench.png)
42
43 Training was performed with [prime-rl](https://github.com/PrimeIntellect-ai/prime-rl) using environments built with the [verifiers](https://github.com/PrimeIntellect-ai/verifiers) library.
44 All training and evaluation environments are available on the [Environments Hub](https://app.primeintellect.ai/dashboard/environments).
45
46 The model, training frameworks, and environments are open-sourced under fully-permissive licenses (MIT and Apache 2.0).
47
48 For more details, see the [technical report](https://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf).
49
50 ## Evaluation
51
52 INTELLECT-3 achieves best-in-class performance on math, coding, and reasoning benchmarks:
53
54 | Benchmark | MATH-500 | AIME24 | AIME25 | LCB | GPQA | HLE | MMLU-Pro |
55 |-----------|----------|---------|---------|--------|------|-----|----------|
56 | INTELLECT-3 | **98.1** | **90.8** | **88.0** | 69.3 | 74.4 | 14.6 | 81.9 |
57 | GLM-4.5-Air | 97.8 | 84.6 | 82.0 | 61.5 | 73.3 | 13.3 | 73.9 |
58 | GLM-4.5 | 97.0 | 85.8 | 83.3 | 64.5 | 77.0 | 14.8 | 83.5 |
59 | DeepSeek R1 0528 | 87.3 | 83.2 | 73.4 | 62.5 | 77.5 | 15.9 | 75.3 |
60 | DeepSeek v3.2 | 96.8 | 88.1 | 84.7 | **71.6** | **81.4** | **17.9** | **84.6** |
61 | GPT-O5S 120B | 96.0 | 75.8 | 77.7 | 69.9 | 70.0 | 10.6 | 67.1 |
62
63 ## Model Variants
64
65 | Model | HuggingFace |
66 |-------|-------------|
67 | INTELLECT-3 | [PrimeIntellect/INTELLECT-3](https://huggingface.co/PrimeIntellect/INTELLECT-3) |
68 | INTELLECT-3-FP8 | [PrimeIntellect/INTELLECT-3-FP8](https://huggingface.co/PrimeIntellect/INTELLECT-3-FP8) |
69
70 ## Serving with vLLM
71
72 The BF16 version can be served on 2x H200s:
73 ```bash
74 vllm serve PrimeIntellect/INTELLECT-3 \
75 --tensor-parallel-size 2 \
76 --enable-auto-tool-choice \
77 --tool-call-parser qwen3_coder \
78 --reasoning-parser deepseek_r1
79 ```
80
81 The FP8 version can be served on a single H200:
82
83 ```bash
84 vllm serve PrimeIntellect/INTELLECT-3-FP8 \
85 --enable-auto-tool-choice \
86 --tool-call-parser qwen3_coder \
87 --reasoning-parser deepseek_r1
88 ```
89
90 ## Citation
91
92 ```bibtex
93 @misc{intellect3,
94 title={INTELLECT-3: Technical Report},
95 author={Prime Intellect Team},
96 year={2025},
97 url={https://huggingface.co/PrimeIntellect/INTELLECT-3}
98 }
99 ```
100