Mistral’s New AI Models Slash Costs and Boost Power
Mistral launches Leanstral and Small 4, powerful open-source AI models cutting costs while boosting coding, reasoning, and enterprise performance.
French AI company Mistral is making a bold move in the open-source space with the release of two powerful models: Leanstral and Mistral Small 4. Together, they signal a shift toward cheaper, more efficient, and highly specialised AI—without sacrificing performance.
From formal mathematical proof systems to enterprise-grade AI workloads, Mistral is positioning itself as a serious challenger in the global AI race.
Leanstral: A Breakthrough in Formal Verification
Leanstral is not your typical AI model—it’s built specifically for formal verification using the Lean 4 proof assistant. That means it can help developers and researchers prove that code works correctly, not just generate it.
What Makes Leanstral Different?
120B total parameters, but only 6B active at once (efficient MoE design).
Designed for proof engineering, not general chat.
Integrates directly with Lean via the Model Context Protocol (MCP).
Available via API, free access (Mistral Vibe), and downloadable weights.
Performance vs Cost: A Huge Gap
Leanstral delivers impressive results at a fraction of the cost of competitors:
Scores 26.3 on FLTEval with two passes.
Costs around $36 per run.
Claude Sonnet 4.6: lower score (23.7) at $549.
Claude Opus 4.6: higher score (39.6) but costs $1,650.
Why This Matters
Formal verification is critical in industries like:
Finance (secure transactions)
Aerospace (safety-critical systems)
Blockchain and smart contracts
Leanstral could dramatically lower the barrier to entry for these use cases.
Mistral Small 4: One Model, Three Roles
Mistral Small 4 is designed as a unified AI system, combining:
Instruction-following
Advanced reasoning
Code generation
Instead of switching between models, users get everything in one.
Key Features
119B parameters with 128 experts, only 6.5B active per token.
Massive 256K token context window.
Handles text and image inputs.
Configurable reasoning depth depending on task.
Performance Gains
40% faster completion times vs Small 3.
3× higher throughput in optimized environments.
Matches or beats GPT-OSS 120B on benchmarks like LiveCodeBench.
Generates 20% less output, improving efficiency.
Deployment Considerations
While efficient in inference, it still requires serious hardware:
Minimum: 4× NVIDIA H100 GPUs for self-hosting.
Also available via API, Hugging Face, and NVIDIA NIM containers.
This makes it ideal for:
AI startups
SaaS platforms
Enterprise automation systems
A Strategic Move: The NVIDIA Alliance
Mistral didn’t just release models—it joined the NVIDIA Nemotron Coalition, a partnership focused on building next-generation open AI systems.
What This Means
Access to NVIDIA DGX Cloud for large-scale training.
Collaboration on multimodal and frontier models.
A stronger push toward open AI ecosystems.
CEO Arthur Mensch summed it up clearly: open models are becoming the foundation of the AI platform economy.
Why This Matters for Businesses and Creators
For anyone building online income streams—whether through content, SaaS, or automation—this shift is important.
Opportunities You Can Leverage
Lower API costs → higher margins.
Open-source flexibility → build custom tools.
Faster models → better UX and engagement.
For example, a content site (like an expat blog or news platform) could:
Auto-generate summaries and updates.
Build AI-driven tools for readers.
Reduce reliance on expensive proprietary APIs.
Enjoyed this? Get the week’s top France stories
One email every Sunday. Unsubscribe anytime.
