Mistral AI Drops Mistral Large 4, a 1T Open-Weight Coding Beast

Mistral AI has released the weights for Mistral Large 4, its flagship multimodal model nicknamed Le Chonk. The company paired the release with Orbital Trace, an interactive 3D map of roughly 19,000 tracked satellites and debris fragments. According to Mistral’s release announcement , the model built the application end to end.

Mistral says the agent retrieved orbital data, planned and wrote the application, opened it in a browser, identified bugs and latency problems, and revised the code in response to an independent reviewer. The demonstration presents a complete software-development loop executed without a human orchestrator.

The model behind the globe

Mistral Large 4 is an open-weight, general-purpose model that handles text and images. Open-weight means developers can download its parameters and run them on their own infrastructure, subject to the license terms.

The model uses a granular Mixture-of-Experts architecture with 1.05 trillion total parameters, 52 billion of which are active for each token. A router sends each token through a subset of specialized components, reducing per-token computation compared with activating the full model. A separate 1.6-billion-parameter vision encoder processes images.

Mistral says the published weights improve on its earlier public preview after additional reinforcement learning, with the largest gains in coding and defensive cybersecurity.

Compute controls

• Low: Simple tasks and lower inference costs. This replaces the previous no-reasoning setting.
• Medium: General development and everyday workloads.
• High: The default mode for complex tasks requiring more inference-time reasoning.

Benchmarks, with the fine print

Mistral reports competitive results across coding and security evaluations. The figures below come from the company and require independent reproduction across different prompts, tools, runtimes, and inference settings.

Mistral also reports performance comparable to GLM-5.3, particularly on cyber tasks, while using 47% fewer tokens per task. Token efficiency can lower API costs and latency, although the practical result depends on reasoning mode, context size, tool calls, and workload.

The loop behind Orbital Trace

Orbital Trace demonstrates an agent workflow that extends beyond code generation. The model translates requirements into a plan, retrieves tracking data, builds the interface, operates the result in a browser, measures performance, and fixes issues until the reviewer’s acceptance criteria pass.

A companion demonstration applies the same approach to scientific software. Mistral Large 4 generates a browser interface that runs a Hartree-Fock calculation, an approximation of electron behavior, as users stretch molecular bonds. The application connects the calculation to a 3D view of the changing molecular orbitals.

Public access to Orbital Trace remains unclear. At least one publication reported finding no standalone application or repository matching the demo. That limits independent inspection of the prompts, generated code, reviewer, test criteria, data-refresh process, and final performance. For now, the project functions as a company showcase rather than a generally available product.

Two deployment paths

The Mistral Large 4 weights are available through Hugging Face for self-hosting with vLLM or SGLang. Hosted access is available through the Mistral API, OpenRouter, Microsoft Foundry, and NVIDIA endpoints , which include an optimized NIM container.

Self-hosters must provision storage for all 1.05 trillion parameters even though 52 billion are active per token. At 16-bit precision, the raw weights would occupy roughly 2.1 TB before runtime overhead. Quantization can reduce that footprint, while production throughput also depends on expert routing, accelerator memory, interconnect bandwidth, batching, and the selected serving engine.

Price and license limits

Mistral lists the following API prices:

The modified MIT license permits use, modification, fine-tuning, and commercialization. Companies offering Mistral Large 4 as a managed Model-as-a-Service must obtain a separate commercial license.

Where Le Chonk fits

Candidate workloads include coding agents that operate across repositories, terminals, and browsers; defensive security systems that reproduce and patch vulnerabilities; and multimodal scientific tools that combine computation with interactive visualization.

Adoption depends on results from internal evaluations using representative repositories, security policies, latency targets, and infrastructure. The release gives teams a self-hostable alternative to closed coding APIs, while Orbital Trace shows the autonomous build-and-test workflow Mistral designed the model to handle.