News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgfnLeanstral 1.5

Leanstral 1.5

An open code agent model built for the Lean 4 proof assistant

Leanstral 1.5 is a mixture-of-experts code agent model from Mistral AI, aimed at work in Lean 4, a proof assistant used for formal mathematics and software specifications. It is an update of the earlier Leanstral model, with 119B parameters, 6.5B active per token, a 256k token context window and text plus image input.

Share on WhatsApp

  • 119B total parameters with 6.5B active per token
  • Mixture of experts: 128 experts, 4 active per token
  • 256k token context window, 200k recommended
  • Text and image input, text output, with tool calling
  • Apache 2.0 licence, also served on the Mistral API

What it is

Leanstral 1.5 is a code agent model built for Lean 4, the proof assistant used to formalise advanced mathematics and to state properties of software such as Rust code. Mistral AI released it as an update to its earlier Leanstral model and places it inside the Mistral Small 4 family. It has 119 billion parameters in total, of which 6.5 billion are active for each token, and it handles a context window of 256k tokens. It reads text and images and writes text. The company positions it as a cheaper alternative to closed-source options for long Lean tasks.

How it came about

The model follows a first release simply called Leanstral, published under the identifier Leanstral-2603. Version 1.5 is described as an updated model of that release and is grouped by Mistral AI with the Mistral Small 4 collection. It is distributed through the Mistral platform and the Hugging Face hub, and the company notes that existing Leanstral installs pick up the new model automatically.

Inside the model

The design is a mixture of experts: 128 experts in total, with 4 used for each token, so 6.5 billion of the 119 billion parameters are active at a time. This keeps serving costs down while holding a large total capacity. The context window reaches 256k tokens, though the company advises staying at or below 200k. Input can be text or images, output is text. A reasoning setting is offered, with 'high' recommended for harder prompts and 'none' for speed. The model can also call tools, for example a helper that runs Lean code and returns results or errors.

What it can do

The model is aimed at Lean 4 work: proving theorems, fixing code in a project, and defining formal objects such as state machine transitions. The company says it can take on long-range tasks that run for hours. Example output shows it revising a proof attempt and proposing a simpler inductive definition to make a lemma easier to prove. It also supports tool calls to compile and run Lean snippets. No benchmark scores are given.

How to use it

Weights are downloadable under the Apache 2.0 licence, with a condition that use must not infringe third-party rights. The model is also served through the Mistral API and the Mistral chat platform, where it can be used on the free plan after enabling Labs models. It runs locally with vLLM 0.24.0 or newer, and works with the Mistral vibe command-line agent and the lean-lsp-mcp tool for Lean.

Questions people ask about Leanstral 1.5

What is Leanstral 1.5 for?

It is a code agent model built for Lean 4, a proof assistant used to formalise mathematics and to state properties of software. Mistral AI says it can prove theorems, fix code in a project and handle tasks that take hours of work.

How large is the model?

It has 119 billion parameters in total. Because it is a mixture of experts with 128 experts and 4 active per token, only 6.5 billion parameters are used for each token, which lowers serving cost.

How do I get it?

The weights are downloadable under the Apache 2.0 licence. You can also reach it through the Mistral API and chat platform on a free plan after enabling Labs models, or run it locally with vLLM and the Mistral vibe agent.

What can it take as input?

It accepts text and images and produces text. It can also call tools, such as a helper that runs or compiles a Lean snippet and returns the result or an error message.

How is it different from the first Leanstral?

Leanstral 1.5 is an updated version of the earlier model released as Leanstral-2603. It belongs to the Mistral Small 4 family and adds multimodal input alongside its code agent features.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Leanstral 1.5 yet, so the facts on this page come from the sources listed above: huggingface.co/mistralai/Leanstral-1.5-119B-A6B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.