INTELLECT-2
INTELLECT-2 starts from QwQ-32B. Community-supplied GPUs contributed to reinforcement learning on verifiable mathematics and coding tasks.
What the release provides
The prime-rl framework runs asynchronous distributed training. The recipe uses GRPO and verifiable rewards, with changes intended to improve stability. The card reports stronger mathematics and coding results but slightly lower instruction-following results than the starting model.INTELLECT-2 official model card
How to use it well
Practical guidance. Allow enough output space for reasoning. Compare success and total output cost across representative tasks before selecting a reasoning budget.
Length instructions are approximate
Prime recommends a 10,000-token reasoning instruction. Other training targets were 2,000, 4,000, 6,000, and 8,000 tokens. The model did not fully learn this control, so an instruction is not a strict output limit.INTELLECT-2 official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Context window | 40,960 tokens Configured position limit. Serving limits depend on the runtime and settings. INTELLECT-2 model configuration |
| Repository license | Apache 2.0INTELLECT-2 official model card |
| Parameters | 32 billionINTELLECT-2 official model card |
| Base model | QwQ-32BINTELLECT-2 official model card |
Related models
- INTELLECT-3
Prime Intellect · INTELLECT
A 106-billion-parameter mixture-of-experts reasoning model with 12 billion active parameters.
Launch date not confirmedDownloadable - INTELLECT-3.1
Prime Intellect · INTELLECT
A further reinforcement-learning release of the INTELLECT-3 reasoning model.
Launch date not confirmedDownloadable