Mistral launches Large 4 open weights for organizations that want to run its trillion-parameter model on their own infrastructure, completing a release that began with a hosted preview on October 6.

 

The October 29 checkpoint adds a practical deployment choice to Mistral’s cybersecurity pitch. Developers can download the model, customize it and serve it without depending entirely on Mistral’s API, although the model’s size and license create important operating limits.

 

The final release adds four material changes:

  • Downloadable FP8 and NVFP4 checkpoints
  • Three reasoning modes for different workloads
  • Lower hosted API prices
  • A separate license condition for model-as-a-service businesses

 

Mistral Launches Large 4 Open Weights After the Preview

Mistral Large 4, also called Le Chonk, is a natively multimodal mixture-of-experts model with about 1.05 trillion total parameters. The system activates approximately 49 billion parameters for each token, or 52 billion when embeddings and output layers are included.

 

The model accepts text and images, generates text and supports a context window of up to one million tokens. Hugging Face lists official FP8 and NVFP4 variants, while community-produced GGUF versions offer additional quantization choices for operators with different memory budgets.

 

This is a material update to the early October preview rather than a new model family. The preview gave developers hosted API access and promised downloadable weights later in the month. The October 29 release fulfills that promise after additional reinforcement learning and testing with government and corporate users.

 

Organizations can now serve the checkpoint with vLLM or SGLang, according to Mistral. Hosted options remain available through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry and Nvidia services, giving customers a choice between direct model operation and managed infrastructure.

 

Large 4 Targets Coding and Defensive Cybersecurity

Mistral positions Large 4 as a model for coding agents, multimodal work and defensive cybersecurity. The company says feedback from security partners helped improve the released checkpoint’s ability to identify vulnerabilities, write patches and support incident-response tasks.

 

Reuters reported that Mistral describes Large 4 as its most capable openly available cybersecurity model. The distinction matters because security teams can inspect, adapt and retain an open-weight system, including during an incident when losing access to an external provider could disrupt response work.

 

Open access also removes a central control point. Once the checkpoint is downloaded, Mistral cannot contain who uses it or guarantee that downstream operators preserve its safeguards. That trade-off is more consequential for a model trained to reason about real software weaknesses than for a general-purpose chatbot.

 

Independent results put Mistral’s performance claims in context. Artificial Analysis scored the released model at 26 on Intelligence Index v5 in its high-reasoning mode, one point above the preview but six points behind the highest-scoring open-weight systems in that assessment.

 

Mistral Cuts API Prices and Adds Reasoning Controls

The hosted version now costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Those rates are roughly half the preview prices, lowering the cost of testing the updated checkpoint before committing to self-hosted infrastructure.

 

Artificial Analysis measured a 63% decline in cost per Intelligence Index task, from $5.54 for the preview to $2.06 for the released model at its high setting. That benchmark is useful for comparison, but it does not measure the hardware, engineering and oversight costs of operating the downloaded weights.

 

Mistral also exposes low, medium and high reasoning modes. Low is intended for lighter workloads, medium for routine tasks and high for harder requests. A development team can therefore reserve deeper reasoning for complex debugging or security analysis instead of paying the maximum inference cost for every prompt.

 

Self-hosting remains a demanding proposition. A one-trillion-parameter checkpoint requires substantial accelerator memory, interconnect bandwidth, storage and operational expertise even when only a fraction of its experts activate for each token. Quantization can reduce the footprint, but it may change quality, throughput and numerical behavior.

 

The Open-Weight License Draws a Commercial Boundary

Large 4 uses the Mistral Ecosystem Open-Weight License, a modified MIT license. Mistral says users may run, modify, fine-tune and commercialize the model, while businesses operating a model-as-a-service offering must obtain a separate commercial license before commercial use.

 

That condition prevents the release from being described simply as MIT-licensed or unrestricted open source. Organizations planning hosted endpoints for third parties need to review the definition of model-as-a-service and any applicable exceptions before building a commercial product around the checkpoint.

 

For enterprises, the decision now shifts from whether the promised weights will arrive to whether independent operation is worth the cost. Large 4 offers control over deployment, customization and continuity, but those benefits come with responsibility for security hardening, monitoring, updates and misuse prevention.

 

Illustrative feature image: Mistral AI co-founder and CEO Arthur Mensch at Techarena 2026, photographed by Jan2342342423 and sourced from Wikimedia Commons. Cropped, resized and converted to WebP; the original is dedicated to the public domain under CC0 1.0.

 

Related Research