Leanstral 2603
Leanstral 2603 adapts the Mistral Small 4 family for Lean 4 proof engineering. Its 119B mixture-of-experts model activates 6.5B parameters per token and accepts text and images.
Proof engineering with a sparse model
The architecture routes each token to four of 128 experts. The card describes integration with Mistral Vibe for work on formal proofs. A proof assistant checks the generated Lean code; the language model’s statement that a proof is correct is not that check.Leanstral 2603 official model card
Advertised and recommended context
The checkpoint supports an advertised 256K context, but the card recommends staying at or below 200K tokens. It also provides reasoning settings of none and high, with high recommended for complex tasks. The March checkpoint remains downloadable after its API retirement.Leanstral 2603 official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | Leanstral-2603Leanstral 119B A6B |
| Context window | 256,000 tokens The model catalog advertises 256k context for this version. This is not a measured recall limit. Leanstral 2603 model and lifecycle record |
| Weight license | Apache 2.0Leanstral 2603 official model card |
| Access and lifecycle | Hosted API retired on 30 June 2026. Downloadable weights remain listed.Leanstral 2603 model and lifecycle recordLeanstral 2603 official model card |
| Parameters | 119B total; 6.5B active per tokenLeanstral 2603 official model card |