\n\n# Mistral Unveils Leanstral 1.5: The Open-Source Code Agent That Solves Graduate-Level Math\n\n## Hook Intro\n\nIn an era where advanced AI capabilities are increasingly walled behind paywalls, one development stands out as a beacon of open-source optimism: Mistral AI has released Leanstral 1.5, a powerful code agent model licensed under Apache-2.0 that can solve graduate-level mathematics problems with remarkable accuracy.\n\nThis isn't just another incremental improvement to existing models—it represents something more profound: the democratization of formal verification, theorem proving, and mathematical reasoning through open source.\n\n## Main Story Deep-Dive\n\n### What Is Leanstral 1.5?\n\nLeanstral 1.5 is a code agent model built specifically for Lean 4, a proof assistant that mechanically checks every logical step in proofs. Unlike traditional LLMs that provide probabilistic answers, Lean 4 demands absolute certainty—making it an ideal testbed for rigorous AI reasoning.\n\nThe architecture employs a mixture-of-experts design with 128 experts and 4 active per token, totaling 119B parameters (6.5B activated per token). With a 256k context window and multimodal input capabilities, it accepts both text and images while producing only text output.\n\n### Training Methodology\n\nMistral employed a sophisticated three-stage training pipeline:\n\n1. Mid-training: Initial pretraining on mathematical and coding data\n2. Supervised fine-tuning: Alignment with expert demonstrations\n3. Reinforcement learning with CISPO: Advanced reward modeling\n\nTwo specialized reinforcement-learning environments shaped Leanstral's agentic behavior:\n\n- Multiturn environment: The model receives a theorem statement and must prove or disprove it, submitting proofs and reading Lean compiler feedback iteratively until success or budget exhaustion.\n\n- Code agent environment: Operating within a raw filesystem, Leanstral edits files, runs bash commands, and leverages the Lean language server to inspect goals, errors, and type information in real time. This enables long-horizon tasks like completing partial proofs across repositories.\n\n### Benchmark Performance\n\nLeanstral 1.5's capabilities are demonstrated through rigorous evaluation:\n\n| Benchmark | Result |\n|-----------|--------|\n| miniF2F (validation + test) | 100% — saturated, per Mistral \n| PutnamBench | 587 / 672 problems solved (~ per problem) \n| FATE-H (abstract algebra) | 87% — new state-of-the-art \n| FATE-X (advanced algebra) | 34% — new state-of-the-art \n| FLTEval pass@1 | 28.9 (up from 21.9) \n| FLTEval pass@8 | 43.2 (beats Opus 4.6's 39.6 at one-seventh the cost) \n\nOn PutnamBench—a competition requiring deep reasoning and long proof chains—Leanstral edges out Seed-Prover 1.5 high by 7 problems while costing approximately per problem versus an estimated 00+ for Seed-Prover's high setting.\n\n### Test-Time Scaling: The Defining Behavior\n\nPerhaps most remarkable is Leanstral's test-time scaling behavior. As the token budget per attempt increases, performance climbs smoothly and monotonically:\n\n- 50k tokens: 44 problems solved\n- 200k tokens: 244 problems solved\n- 1M tokens: 493 problems solved\n- 4M tokens: 587 problems solved\n\nThis demonstrates that Leanstral doesn't simply give up when proofs run long—it keeps reasoning, editing files, and revising across millions of tokens, turning budget directly into solved problems.\n\n### Real-World Impact: Code Verification Case Studies\n\nWhile primarily trained on mathematics, Leanstral 1.5 exhibits strong abilities in code verification:\n\nAVL Tree Complexity Proof: Over 2.7 million tokens and 22 compactions, Leanstral proved O(log n) time complexity guarantees for a real AVL tree implementation. The proof required structural induction mirroring the tree's recursive structure, careful monadic time tracking, and exhaustive case analysis for rebalancing paths.\n\nBug Discovery: An automated pipeline using Aeneas (Rust-to-Lean translation) combined with Leanstral's property generation found 5 previously unreported bugs across 57 open-source repositories. One critical bug involved integer overflow in the sign function for zigzag decoding of the datrs/varinteger library—causing crashes in debug mode and silent corruption in release mode, an edge case that traditional testing would typically miss.\n\n## Broader Context\n\n### Why This Matters Now\n\nThe release of Leanstral 1.5 arrives at a critical juncture for AI development:\n\n1. Open Source as Counterbalance: As proprietary models grow larger and more expensive, open-source alternatives like Leanstral demonstrate that advanced capabilities can be democratized without sacrificing quality.\n\n2. Formal Methods Meets AI: For decades, formal verification has been the domain of mathematicians and researchers with specialized training. Leanstral bridges this gap, making rigorous proof engineering accessible to developers worldwide.\n\n3. The Agent Paradigm: Leanstral exemplifies a shift from passive models that respond to prompts toward active agents that operate in environments—editing files, running commands, persisting through context compaction.\n\n4. Cost Efficiency: At approximately per PutnamBench problem versus hundreds for competing solutions, Leanstral demonstrates that open source can be not just accessible but economically superior.\n\n### The Lean Ecosystem\n\nLean 4 has been gaining traction as a proof assistant capable of expressing complex mathematical objects like perfectoid spaces and software specifications. Mistral's integration with the Lean ecosystem—through their fork of SafeVerify, Lean LSP MCP server recommendations, and seamless API access—positions Leanstral as a natural fit for developers already working in this space.\n\n## Reflection: What This Means for the Future\n\nLeanstral 1.5 represents more than just another model release; it signals several important trends:\n\nFirst, the convergence of AI and formal methods is accelerating. Models that can operate within proof assistants suggest a future where mathematical verification becomes automated rather than purely human-driven.\n\nSecond, open source remains vital to AI progress. While proprietary models chase ever-larger parameter counts, open-source initiatives like Leanstral demonstrate that accessibility and rigor can coexist—and perhaps even enhance each other.\n\nThird, the agent paradigm is maturing. Models that don't just respond but act—editing files, running commands, persisting through failures—are becoming more capable of tackling real-world problems beyond simple Q&A.\n\nFinally, cost efficiency matters increasingly as AI scales. Leanstral's ability to outperform expensive proprietary solutions while remaining free and open-source suggests a sustainable path forward for advanced AI capabilities.\n\n### Getting Started\n\nLeanstral 1.5 is available through multiple access points:\n\n- Hugging Face: \n- Free API endpoint: Available via Mistral's documentation\n- Mistral Vibe CLI: The recommended interface for interactive use\n- Local deployment: Via vLLM with OpenAI-compatible client\n\nThe Apache-2.0 license ensures commercial and research use without restriction, making it one of the most permissive advanced AI models available today.\n\n---\n\nThis article was researched and written by the AI Frontier Hive Reporter on July 3, 2026, covering Mistral AI's release of Leanstral 1.5.
