Mistral AI Launches Leanstral and Small 4 at GTC 2026 — First Open-Source Formal Verification Coding Agent Beats Claude at 1/92nd the Cost

Mistral AI made two major releases at NVIDIA GTC 2026 on March 16, 2026, establishing itself as a leading force in efficient, open-source agentic AI.

LEANSTRAL — FORMAL VERIFICATION CODING AGENT:

Leanstral is the first open-source coding agent specifically designed for Lean 4 formal verification. Unlike wrapper agents around generalist models, Leanstral is purpose-built for proof engineering.

Key characteristics:

• 119B total parameters, Mixture-of-Experts architecture with only 6B active per token
• Apache 2.0 license, fully open-source
• Trained specifically for realistic formal repository operations, not just isolated math problems
• Supports MCP integration (trained specifically with lean-lsp-mcp)
• Available via Mistral Vibe agent scaffold, free API, and downloadable weights on HuggingFace

Benchmark results on FLTEval (formal proof engineering):

• Leanstral pass@1: 21.9 ($18) — competitive with Claude Haiku (23.0 at $184)
• Leanstral pass@2: 26.3 ($36) — beats Claude Sonnet (23.7 at $549)
• Leanstral pass@16: 31.9 ($290) — approaches Claude Opus (39.6 at $1,650)
• Outperforms GLM5-744B and Kimi-K2.5-1T despite being 6-165x smaller in active parameters

The cost efficiency is extraordinary: Leanstral delivers comparable quality to Sonnet at 1/15th the cost, and approaches Opus quality at 1/6th the cost.

MISTRAL SMALL 4 — UNIFIED REASONING MODEL:

Simultaneously, Mistral released Small 4, a new flagship efficient model:

• 119B parameters total, 128 expert modules, only 4 active per query (6B active)
• Unifies text, reasoning, and vision in a single model
• Controllable reasoning depth — users choose fast response vs. deep thinking
• 40% faster than predecessor, 3x more queries per second
• Apache 2.0 license
• Available on HuggingFace, Mistral API, and NVIDIA platforms
• Matches or exceeds specialized Magistral models on internal benchmarks

NVIDIA NEMOTRON COALITION:

Mistral joined NVIDIA Nemotron Coalition, a global initiative to advance open AI models through shared expertise, data, and compute. This positions Mistral alongside LangChain and other partners in NVIDIA's enterprise agentic ecosystem.

The release trended at #5 on Hacker News with 614 points and 132 comments, indicating strong developer interest.

The formal verification angle is strategically significant: as AI-generated code reaches 42% of all commits (heading to 65% by 2027 per industry estimates), the bottleneck shifts from code generation to code verification. Leanstral addresses this directly.