{"id":101872,"date":"2026-07-04T18:50:17","date_gmt":"2026-07-04T18:50:17","guid":{"rendered":"https:\/\/youzum.net\/mistral-ai-releases-leanstral-1-5-an-apache-2-0-lean-4-code-agent-model-solving-587-of-672-putnambench-problems\/"},"modified":"2026-07-04T18:50:17","modified_gmt":"2026-07-04T18:50:17","slug":"mistral-ai-releases-leanstral-1-5-an-apache-2-0-lean-4-code-agent-model-solving-587-of-672-putnambench-problems","status":"publish","type":"post","link":"https:\/\/youzum.net\/fr\/mistral-ai-releases-leanstral-1-5-an-apache-2-0-lean-4-code-agent-model-solving-587-of-672-putnambench-problems\/","title":{"rendered":"Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems"},"content":{"rendered":"

Today, Mistral AI released Leanstral 1.5<\/strong>. It is a code agent model built for Lean 4. The release targets automated theorem proving and proof engineering. Weights are open under Apache 2.0. A free API endpoint, leanstral-1-5<\/code>, is now live.<\/p>\n

Leanstral 1.5 updates the earlier Leanstral-2603 model. It belongs to the Mistral Small 4 family. <\/p>\n

What is Leanstral 1.5<\/strong><\/h2>\n

Leanstral 1.5 is a code agent model for Lean 4<\/a>, a proof assistant. A proof assistant checks every logical step mechanically. Lean 4 can express objects like perfectoid spaces and properties of Rust fragments.<\/p>\n

The architecture is a mixture-of-experts, or MoE. An MoE routes each token to a few specialized sub-networks. This keeps compute low while total capacity stays large. Leanstral uses 128 experts, with 4 active per token.<\/p>\n

Total size is 119B parameters, with 6.5B activated per token. Context length is 256k tokens. Input is multimodal, accepting text and image. Output is text only.<\/p>\n

How Mistral Trained Leanstral 1.5<\/strong><\/h2>\n

Training runs in three stages. These are mid-training, supervised fine-tuning, then reinforcement learning with CISPO. Two reinforcement-learning environments shaped the model\u2019s agentic behavior.<\/p>\n

In the multiturn environment<\/strong>, the model receives a theorem statement. It must prove or disprove it. It submits a proof, then reads Lean compiler feedback. It refines across attempts until it succeeds or exhausts its budget.<\/p>\n

In the code agent environment<\/strong>, Leanstral works inside a raw filesystem. It edits files, runs bash commands, and uses the Lean language server. That server exposes goals, errors, and type information in real time.<\/p>\n

This lets it complete partial proofs, build auxiliary lemmas, and persist through context compaction. Compaction compresses earlier context so long tasks still fit the window. Correctness is verified by Mistral\u2019s fork of SafeVerify against target theorems.<\/p>\n

Benchmarks and Performance<\/strong><\/h2>\n

Mistral team reports that Leanstral 1.5 saturates miniF2F. It reaches 100% on both the validation and test sets. It solves 587 of 672 PutnamBench problems.<\/p>\n

The model sets a new state-of-the-art on the FATE-H and FATE-X algebra benchmarks. Mistral lists 87% on FATE-H and 34% on FATE-X. On FLTEval, pass@1 rises from 21.9 to 28.9. Pass@8 rises from 31.9 to 43.2.<\/p>\n

FLTEval is built from real pull requests to the Fermat\u2019s Last Theorem repository. On it, Leanstral surpasses Opus 4.6\u2019s 39.6 at one-seventh the cost. It also widens its lead over open-source models three to ten times larger. Pass@8 means eight attempts are allowed per problem.<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nLeanstral 1.5<\/th>\nDetail<\/th>\n<\/tr>\n<\/thead>\n
miniF2F (val + test)<\/td>\n100%<\/td>\nSaturated, per Mistral<\/td>\n<\/tr>\n
PutnamBench<\/td>\n587 \/ 672<\/td>\n~$4 per problem<\/td>\n<\/tr>\n
FATE-H<\/td>\n87%<\/td>\nNew state-of-the-art<\/td>\n<\/tr>\n
FATE-X<\/td>\n34%<\/td>\nNew state-of-the-art<\/td>\n<\/tr>\n
FLTEval pass@1<\/td>\n28.9<\/td>\nUp from 21.9<\/td>\n<\/tr>\n
FLTEval pass@8<\/td>\n43.2<\/td>\nBeats Opus 4.6\u2019s 39.6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

On PutnamBench, Leanstral edges Seed-Prover 1.5 high by 7 problems. It does so at about $4 per problem. Mistral estimates Seed-Prover\u2019s high setting near $300 or more per problem.<\/p>\n

That setting runs a budget of 10 H20-days per problem. Mistral also compares against Goedel-Architect and AxProverBase. It notes Aleph Prover costs roughly $54 to $68 per problem.<\/p>\n

Test-time scaling is the model\u2019s defining behavior. Raising the token budget per attempt lifts PutnamBench Pass@8. Mistral team reports 44 solved at 50k, 244 at 200k, 493 at 1M, and 587 at 4M. The interactive explorer below lets you scrub across that same curve.<\/p>\n

<\/p>\n

\n

Case Studies and Use Cases<\/strong><\/h2>\n<\/p>

Leanstral trained mainly on mathematics, but it also verifies code. Mistral team documents two case studies that matter for engineers.<\/p>\n