Mistral releases Leanstral 1.5: 119B parameter open-source model for Lean 4 proof assistance
Mistral AI has released Leanstral 1.5, an open-source 119B parameter mixture-of-experts model designed specifically for Lean 4 proof assistance. The model features 128 experts with 4 active per token (6.5B activated parameters), a 256k token context window, and multimodal input capabilities.
Mistral releases Leanstral 1.5: 119B parameter open-source model for Lean 4 proof assistance
Mistral AI has released Leanstral 1.5, an open-source 119B parameter mixture-of-experts model designed specifically for Lean 4, a proof assistant used for expressing complex mathematical objects and software specifications.
Technical specifications
Leanstral 1.5 is built as part of the Mistral Small 4 family and uses a mixture-of-experts architecture:
- Total parameters: 119B
- Active parameters per token: 6.5B (128 experts, 4 active)
- Context window: 256k tokens (200k recommended)
- Input: Text and images (multimodal)
- Output: Text only
The model supports a "reasoning effort" parameter with two settings: 'none' for standard operation and 'high' for chain-of-thought reasoning on complex prompts. Mistral recommends temperature 1.0 for inference.
Deployment and access
Leanstral 1.5 is available free through Mistral's API with no requirement for a Pro subscription. Users must opt into "Enable Labs models" in their Mistral account settings to access the model.
The model integrates with Mistral's Vibe CLI tool as a code agent, accessible via the command vibe --agent lean. According to Mistral, "Leanstral is capable of doing very long range tasks which requires hours of work."
For local deployment, Mistral recommends vLLM with specific configuration:
- Minimum vLLM version: 0.24.0
- Recommended tensor parallel size: 4
- Maximum model length: 200,000 tokens
- Attention backend: FLASH_ATTN_MLA
Integration with Lean ecosystem
The model works with the lean-lsp-mcp, a Model Context Protocol tool for AI agents to interact with Lean's language server. This enables the model to directly interact with Lean projects, prove theorems, and fix code within the Lean 4 environment.
Mistral recommends running the Vibe CLI inside VS Code's terminal and within Lean project directories for optimal workflow integration.
What this means
Leanstral 1.5 represents a specialized application of large language models for formal verification and mathematical proof assistance. The free API access removes cost barriers for researchers and developers working with Lean 4, while the mixture-of-experts architecture keeps inference costs manageable. The 256k context window enables the model to work with large codebases and complex proofs that require extensive context. This release positions Mistral as a key player in AI-assisted formal mathematics, an area that has traditionally relied on closed-source models or general-purpose coding assistants not optimized for proof systems.
Related Articles
Reflection unveils 501B-parameter Beam, Mistral previews 1T-parameter Large 4, both open-weight
Reflection introduced Beam, a 501B-parameter mixture-of-experts model with 23B active parameters. Mistral said it is finishing Mistral Large 4, a 1T-parameter multimodal model with 49B active parameters. Both companies plan open-weight releases in October, and both are positioning the models against Chinese open-weight leaders.
Google releases EmbeddingGemma 2, a 740M-parameter open multimodal embedding model scoring 78.68 on MTEB Code
Google released EmbeddingGemma 2, an open 740-million-parameter model that embeds text, images, video, audio, and code. It scores 78.68 on MTEB (Code), up from 68.76 for its predecessor, and Google claims it beats models up to twice its size on multimodal embedding benchmarks.
Google DeepMind releases EmbeddingGemma 2: 740M-parameter open embedding model spanning text, image, video, audio
Google DeepMind has released EmbeddingGemma 2, an Apache 2.0 open embedding model with 740M total parameters that maps text, images, video and audio into a single 768-dimensional vector space. According to the model card, it improves code retrieval on MTEB (code, v1) from 68.76 to 78.68 over its predecessor while keeping an 8,192-token context window.
Google releases EmbeddingGemma 2: 740M-parameter multimodal embedding model under Apache 2.0
Google announced EmbeddingGemma 2, a 740M-parameter natively multimodal embedding model built on the Gemma 4 architecture and released under Apache 2.0. Google says it runs in ~191MB of active RAM for text-only weights and ~567MB for the full multimodal model on a quantized Pixel 11 Pro. Google also released a Mac app, AI Edge Foresight, to demonstrate it.
Comments
Loading...