Leanstral: 6B Active Parameters Model Sets New SOTA in Formal Math Proofs
Bam4d · x · 2026-08-07
Leanstral is a series of generalist code-agent models for Lean 4, featuring 119B total but only 6B active parameters. Operating directly within the open-sourced Mistral Vibe code-agent harness without specialized prover scaffolding, it scales performance smoothly via context compaction and increased test-time compute.
Despite its size, the model rivals larger proprietary systems: it saturates miniF2F, solves 587/672 problems on PutnamBench, and achieves a new state-of-the-art 34% on FATE-X and 43.2% on FLTEval. Furthermore, an automated pipeline built on Leanstral uncovered previously unknown bugs in real-world open-source repositories. The model is open-sourced under Apache-2.0.
Anecdote: The author noted on X that the paper was previously rejected by arXiv for "not having enough scientific contributions" before being published on alphaXiv.
Related event: Leanstral Model Released, Setting New SOTA in Theorem Proving(2 posts)→
More from coding & agent
- DeepMind turns coding agent Antigravity into a science tool, surfacing novel antimicrobial peptide in one session — pushmeet · 2026-10-16
- Cross-review coding workflow: Claude writes the plan, Codex audits it — theteknosaur · 2026-10-16
- Simon Willison: code mode is the future of tool calling, if it can be sandboxed — springrod · 2026-10-16
- Normalizing Claude's schema-free MCP output before it hits downstream n8n nodes — snowingbol · 2026-10-16
- Agent named 'Ilya' reportedly scores 1.1pt higher on benchmarks — daniel_mac8 · 2026-10-16
- A plain notes.md fixed my agent's cross-session amnesia better than chat history — rahul_watertech · 2026-10-16