Mistral AI’s latest lab release, Leanstral 1.5, is positioned as a dedicated proof‑engineering engine for the Lean 4 language. It combines a 119 billion‑parameter mixture‑of‑experts (MoE) backbone with 6.5 billion active parameters per token, delivering the compute density of a large language model while keeping inference latency comparable to smaller decoders. The MoE routing architecture activates only the most relevant expert sub‑networks for a given token, which explains the low active‑parameter count and the model’s ability to handle a 256 k token context window—a size crucial for complex formal proof scripts that often span many lemmas and definitions[1].

Technical trade‑offs

  • Routing latency – MoE models incur an extra routing step. In practice, the added microsecond‑scale overhead can accumulate for very long contexts, potentially slowing batch proof searches that rely on near‑real‑time feedback.
  • Specialisation vs. generalisation – Leanstral 1.5 is fine‑tuned on Lean 4 theorem‑proving corpora. While this yields strong auto‑formalisation performance, the model shows reduced fluency on non‑technical prose and may generate malformed Lean code when prompted outside its domain.
  • Bias and hallucination – The training data includes public Lean repositories and academic papers. Biases inherit from those sources (e.g., over‑reliance on classical logic patterns, under‑representation of constructive proofs). Users have reported occasional “proof‑by‑assertion” outputs that appear valid syntactically but lack a supporting proof term, a classic hallucination risk in code‑generation models[2].

Cost and deployment implications

Leanstral 1.5 is listed at $0 per‑token on Mistral’s Labs tier, accessible through the web playground or API without a commercial license. For enterprises, this removes direct inference spend but shifts cost considerations to infrastructure: the 256 k context requires GPUs with ≥40 GB VRAM to store activation maps for the active experts, and external storage for the full 119 B parameter checkpoint (≈ 460 GB). Mistral offers cloud deployment via its Compute product and a self‑deployment option that lets firms host the model behind corporate firewalls, a critical feature for IP‑sensitive proof work[1].

Organizational impact

  • Accelerated verification pipelines – By auto‑formalising informal specifications into Lean code, Leanstral 1.5 can shrink the human‑in‑the‑loop verification cycle from weeks to days, improving time‑to‑market for safety‑critical software.
  • Skill shift – Teams will need proof‑engineers familiar with Lean 4 and prompt‑engineering for theorem provers. Mistral’s documentation includes a model‑selection guide and prompt‑engineering best practices, but enterprises must invest in training or hire niche talent.
  • Risk management – The model’s hallucination profile demands a verification layer that re‑runs generated proofs through a trusted Lean kernel. Automated guardrails (e.g., the built‑in moderations API) can filter unsafe outputs, but they do not replace formal proof checking.

When Leanstral 1.5 may struggle

  • Heavy constructive mathematics – The model has been observed to prefer classical reasoning patterns, making it less efficient on proofs that require explicit constructivity.
  • Cross‑language translation – Attempting to convert proofs from Coq or Isabelle into Lean via Leanstral 1.5 often yields syntactic mismatches, reflecting its narrow fine‑tuning.
  • Very large proof trees – Although the 256 k context is generous, proof scripts that exceed this limit must be chunked, which can break the model’s internal state and reduce accuracy. Enterprises evaluating Leanstral 1.5 should pilot a dual‑pipeline: use the model for rapid proof sketching, then feed the sketches into a deterministic Lean verifier. This approach captures the