INTELLECT-3
Open 106B MoE reasoning model from Prime Intellect
Prime Intellect released INTELLECT-3, a 106 billion parameter mixture of experts model focused on reasoning. It was built by fine-tuning GLM-4.5-Air-Base with reinforcement learning and is available under permissive open licences.
- 106 billion parameters with 12 billion active per token
- Trained using reinforcement learning and verifiers
- Scores 98.1 on MATH-500 benchmark
- Released under MIT and Apache 2.0 licences
- Based on GLM-4.5-Air-Base architecture
What it is
INTELLECT-3 is a large language model designed for complex reasoning tasks. Developed by Prime Intellect, it contains 106 billion parameters in total, with 12 billion active at any given time due to its mixture of experts architecture. The model specialises in mathematics, coding, and logical deduction. It serves as an open-weight alternative for developers seeking high-performance reasoning capabilities without restrictive licensing barriers.
How it came about
The model was derived from the GLM-4.5-Air-Base checkpoint. Prime Intellect applied supervised fine-tuning followed by extensive reinforcement learning to enhance its reasoning abilities. The training process utilised the prime-rl framework and custom verifiers library. All environments used for training and evaluation were made publicly available on the Environments Hub platform alongside the model release.
Inside the model
This system uses a mixture of experts design, meaning only a subset of its neural network activates for each token. This allows a total size of 106 billion parameters while keeping computational costs lower during inference. The training pipeline relied on reinforcement learning with verified outputs to improve accuracy in math and code generation. The base weights come from the GLM-4.5-Air-Base model, which was then adapted for specific reasoning tasks.
What it can do
Benchmark results show strong performance in mathematical and coding challenges. On the MATH-500 test, it scored 98.1, and on AIME24, it reached 90.8. It achieved 69.3 on the LCB coding benchmark and 74.4 on GPQA. While it trails some larger proprietary models on general knowledge tests like MMLU-Pro, where it scored 81.9, it excels in specialised reasoning domains compared to many open alternatives.
How to use it
The model weights are fully open under MIT and Apache 2.0 licences. Two variants are provided: a standard BF16 version and an FP8 quantised version for efficiency. The standard model can run on two H200 graphics cards, while the FP8 version fits on a single H200. Serving instructions are provided for vLLM, including tool call parsing support.
Questions people ask about INTELLECT-3
What is the architecture of INTELLECT-3?
It uses a mixture of experts design with 106 billion total parameters. Only 12 billion parameters are active for each token processed, making it more efficient to run than dense models of similar total size.
How was INTELLECT-3 trained?
The model started from GLM-4.5-Air-Base. It underwent supervised fine-tuning and then large-scale reinforcement learning using the prime-rl framework and verifiers library to improve reasoning skills.
Can I use INTELLECT-3 commercially?
Yes. The model weights and training environments are released under fully permissive MIT and Apache 2.0 licences, allowing for broad commercial and research use without restrictive conditions.
What hardware is needed to run it?
The standard BF16 version requires two H200 GPUs. A quantised FP8 version is also available, which can be served on a single H200 GPU using vLLM software.
Good reading and sources
There is no English Wikipedia article on INTELLECT-3 yet, so the facts on this page come from the sources listed above: huggingface.co/PrimeIntellect/INTELLECT-3. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.