Leanstral: Open-Source Code Agent for Lean 4 and Formal Proof Engineering

What Leanstral Is
Leanstral is an open-source code agent specifically designed for Lean 4, a proof assistant capable of expressing complex mathematical objects and software specifications. Unlike existing proving systems that act as wrappers around large generalist models, Leanstral is trained for operating in realistic formal repositories with 6B active parameters.
Key Technical Details
The model uses a highly sparse architecture optimized for proof engineering tasks. It leverages parallel inference with Lean as a verifier, making it both performant and cost-efficient. Leanstral supports arbitrary MCPs through Mistral Vibe and was specifically trained to achieve maximal performance with the frequently used lean-lsp-mcp.
Performance Benchmarks
Leanstral was evaluated using FLTEval, a new evaluation suite focused on realistic proof engineering scenarios rather than isolated mathematical problems. The benchmarks compare completion of formal proofs and correct definition of new mathematical concepts in PRs to the FLT project.
Against Open-Source Models
- Leanstral-120B-A6B achieves a score of 26.3 with pass@2 (2 inference passes)
- GLM5-744B-A40B caps at approximately 16.6
- Kimi-K2.5-1T-32B caps at approximately 20.1
- Qwen3.5-397B-A17B requires 4 passes to reach 25.4
- Leanstral scales linearly, reaching 29.3 at pass@4 and 31.9 at pass@16
Against Claude Family
- Leanstral pass@2 (score 26.3) beats Sonnet (23.7) by 2.6 points
- Cost: Leanstral $36 vs. Sonnet $549
- Leanstral pass@16 reaches 31.9, beating Sonnet by 8 points
- Claude Opus 4.6 leads with 39.6 but costs $1,650 (92× Leanstral's cost)
- Haiku scores 23.0 at $184
Case Study Example
When presented with a real-world question from Proof Assistants Stack Exchange about a script that stopped compiling in Lean 4.29.0-rc6, Leanstral successfully built test code to recreate the failing environment. It diagnosed that a def T2 := List Bool was blocking the rw tactic from matching patterns due to definitional equality issues. The fix proposed was swapping def for abbrev since abbrev creates a transparent alias.
Availability
Leanstral weights are released under Apache 2.0 license, available in agent mode within Mistral Vibe, and through a free API endpoint. A tech report detailing the training approach will also be released.
📖 Read the full source: HN AI Agents
👀 See Also

Traversable Skill Graph for Persistent AI Agent Memory in Codebases
A developer built a three-layer skill graph system that lives inside a codebase, enabling AI coding assistants to maintain persistent memory across sessions. The system uses progressive disclosure with self-directing instructions instead of monolithic context files.

Open-source solo RPG engine uses three Claude instances for parsing, narration, and direction
EdgeTales is an open-source text-based solo RPG engine where dice mechanics determine outcomes and Claude AI generates atmospheric prose. The system uses three Claude instances in a pipeline: Brain (Haiku) for parsing input to JSON, Narrator (Sonnet) for writing prose, and Director (Haiku) for async scene analysis.

Founder Operations in Claude: 19 Reusable Skills for Early-Stage Startups
A founder who exited their first startup published 19 Claude-compatible skill prompts for functions like positioning, pricing, prospecting, and copy — based on their own SOPs and Notion workflows.
How I Feel About AI: A Developer's Honest Take on Reasoning, Doom, and the Open Web
A developer reflects on AI: gut-feeling LLMs with emergent reasoning, the threat of superintelligence, and the toll on the open web and artists.