Mistral releases Leanstral 1.5 for proof and code checking

Mistral released Leanstral 1.5, a free model under the . It is presented as a 119B-scale model with 6B , meaning only part of the model is used for each run.

Its main strength is , where software, code rules, or mathematical claims are checked with strict logic. It reached the limit of the miniF2F benchmark, solved 587 of 672 PutnamBench problems, and scored 87% on FATE-H and 34% on FATE-X.

Training used mid-training, , and with CISPO. It is meant for and formal proof engineering, and in real code checks it found 5 previously unknown bugs across 57 tested .

Key points

  • Leanstral 1.5 is available under the .
  • It is described as a 119B-scale model with 6B .
  • It solved 587 out of 672 PutnamBench problems.
  • It scored 87% on FATE-H and 34% on FATE-X.
  • It found 5 previously unknown bugs in 57 tested code .
Read original