Mistral's tiny AI beats models 100x its size at math
Leanstral uses 6B parameters but outperforms 1-trillion-parameter models at math proofs. 92x cheaper than Claude Opus. Free to try.
Mistral AI just proved something important about the future of AI: bigger isn't always better. Their new model, Leanstral, uses only 6 billion active parameters (a measure of a model's size) but outperforms models that are 100 to 170 times larger at a very specific task — writing mathematical proofs.
The announcement exploded on Hacker News with 763 points, making it one of the most-discussed AI stories of the day.
David vs. Goliath — By the Numbers
Leanstral was tested on FLTEval, a benchmark (standardized test) that measures how well AI can complete real mathematical proofs in actual code repositories — not just solve textbook problems.
Performance Scores (higher = better)
🏆 Leanstral (6B active parameters): 26.3 — cost: $36
• Qwen3.5 (397B parameters): 25.4 — needed double the attempts
• Kimi-K2.5 (1 trillion parameters): 20.1
• GLM5 (744B parameters): 16.6
• Claude Sonnet: 23.7 — cost: $549
• Claude Opus: 39.6 — cost: $1,650
The takeaway: Leanstral beats Claude Sonnet while costing 15x less. Claude Opus still wins on raw quality, but it costs 92x more for this specific task.
Why This Matters Beyond Math
You might think: "I don't write mathematical proofs, why should I care?" Here's why this story is bigger than math:
The era of specialist AI is here. Instead of one giant AI doing everything, we're seeing small, focused models that dominate their niche. Just like you'd hire a specialist accountant over a generalist for tax law, specialized AI models can outperform massive ones at specific tasks.
Cost is dropping dramatically. If a 6B-parameter model can match or beat models hundreds of times larger, the cost of AI-powered specialized work is about to plummet. Today it's math proofs; tomorrow it could be legal document review, medical coding, or financial auditing.
Small models = local models. At 6B parameters, Leanstral is small enough to run on your own computer. No cloud, no API costs, no sending sensitive data to external servers.
What Leanstral Actually Does
For the technically curious: Leanstral works with Lean 4, a programming language used by mathematicians and computer scientists to write formal proofs — code that mathematically guarantees something is true. Think of it as a spell-checker, but for logic instead of grammar.
The model can:
- Complete complex mathematical proofs in real-world projects
- Diagnose compilation errors and find root causes
- Translate proofs between different formal languages
- Explain its reasoning step by step
Try It Free
Mistral is offering Leanstral through a free (or near-free) API for a limited time:
- Mistral Vibe — Use the
/leanstralcommand for zero-setup access - Labs API — Free endpoint:
labs-leanstral-2603 - Self-hosted — Download the Apache 2.0 licensed weights (completely free, open-source) and run it yourself
Sign up at the Mistral Console or read the full announcement.
Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News
Stay updated on AI news
Simple explanations of the latest AI developments