HN
Today

Mistral Releases Leanstral

Mistral has unleashed Leanstral, an open-source AI agent purpose-built for formal proof engineering in Lean 4, aiming to tackle the bottleneck of human verification for AI-generated code in high-stakes domains. This specialized model promises cost-efficiency over larger generalist LLMs, though its raw performance still trails top-tier proprietary models. The Hacker News community is debating its comparative value, weighing its open nature and niche focus against the higher performance of more expensive alternatives.

130
Score
23
Comments
#1
Highest Rank
22h
on Front Page
First Seen
Mar 16, 10:00 PM
Last Seen
Mar 17, 7:00 PM
Rank Over Time
112222245411355691214141522

The Lowdown

Mistral AI has introduced Leanstral, a new open-source AI code agent specifically designed for formal proof engineering within the Lean 4 proof assistant. This development targets the critical challenge of manually verifying AI-generated code in high-stakes applications, from advanced mathematics to mission-critical software.

  • Purpose-Built for Lean 4: Leanstral is the first open-source code agent trained explicitly for Lean 4, a proof assistant capable of complex mathematical expressions and software specifications.
  • Open and Accessible: The model's weights are released under an Apache 2.0 license, available through Mistral Vibe, and via a free API endpoint. A technical report and a new evaluation suite, FLTEval, are also planned.
  • Efficient Architecture: Featuring 6B active parameters and a sparse architecture, Leanstral is optimized for proof engineering tasks. It leverages parallel inference with Lean as a verifier, promising performance and cost-efficiency.
  • Benchmarking: On its new FLTEval suite, Leanstral demonstrates significant efficiency gains over larger open-source models (like Qwen, Kimi, GLM5) and offers competitive performance against closed-source models (Claude Sonnet) at a fraction of the cost, though it does not yet surpass Claude Opus in raw score.
  • Real-World Application: Case studies highlight Leanstral's ability to diagnose and propose fixes for breaking changes in Lean 4 and translate program definitions from Rocq into Lean, including proving properties.

Leanstral represents a significant step towards a future where AI agents not only generate code but also formally prove its correctness, offering an accessible and cost-effective tool for the formal verification community.

The Gossip

Performance and Price Point Ponderings

The HN community is keenly analyzing Leanstral's benchmark results, particularly its cost-to-performance ratio against top-tier models like Claude Opus. While Leanstral is significantly cheaper, its lower raw score sparks debate on whether the cost savings justify a performance hit for high-stakes formal verification tasks. Some users argue Opus's higher cost is warranted for superior accuracy, while others point to Leanstral's efficiency, especially with multiple passes, as a compelling advantage.

Passes and Parallel Processing Possibilities

Commenters expressed curiosity about the 'passes' metric used in Leanstral's evaluation, which quantifies multiple attempts at solving a problem. This led to discussions about advanced strategies for boosting performance, such as 'LLM alloys' where different models are used sequentially or in parallel for different passes, leveraging their unique strengths to achieve higher accuracy than any single model alone.

Mistral's Model Momentum

Beyond Leanstral's specific capabilities, the discussion touched on general sentiment towards Mistral AI. Many users praised Mistral for their consistent releases and preferred their models for various daily tasks, citing reliability and output quality. However, a cynical note also questioned the company's business viability, balanced by a humorous counterpoint regarding taxpayer funding.