Mistral Large 4 Goes Open Weight: Self Host It or Use the API?

If your team wanted a frontier sized model it could keep inside its own firewall, Mistral just made that possible. On October 29, the Paris company released the open weights of Mistral Large 4, the trillion parameter model it nicknames "Le Chonk," three weeks after a preview version first appeared.

That leaves developers and IT teams with a real choice: download the weights and run Mistral Large 4 on their own hardware, or keep paying per token through Mistral's API and cloud partners. Here is how the two paths compare.

What Mistral actually released

Mistral Large 4 is a mixture of experts model with about 1 trillion total parameters. Only a slice is active for each token: 49 billion, or 52 billion counting embeddings and output layers, according to the model card on Hugging Face. The model routes each token through 4 of 128 experts, plus one shared expert.

The rest of the spec sheet:

• Context window: 1 million tokens
• Inputs: text and images, with text output
• Reasoning modes: low, medium and high
• Training: 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers, according to SiliconANGLE

Mistral calls it the "best open weight model for coding" in its release announcement. Its own numbers back the claim only partly. On the Artificial Analysis Coding Agent Index, Large 4 scored 48%, which Mistral describes as effectively tied with GLM-5.3, while using 47% fewer tokens. On Terminal-Bench 4.0 it rose from 28% in the preview to 41%. These are vendor reported results, and Mistral says a full technical report is still coming.

Option one: download the weights and host Mistral Large 4 yourself

What you get

Control is the main draw. Because the weights are public, a company can deploy the model on its own servers, behind its own firewall, without sending prompts or code to an outside provider. That matters for banks, government agencies and any team handling source code it cannot share. It also removes the risk that a hosted model changes under you, a problem we saw when Google's Gemini 4 Carbon shipped on a two week release cycle.

The license is a modified MIT license called the Mistral Ecosystem Open Weight License. Using and commercializing the model is free, with one carve out: a business that sells the model itself as a service needs a separate commercial license from Mistral.

What it costs

This is not a laptop model. The example setups on the model card spread it across a node of 8 GPUs, and a compressed NVFP4 version targets Nvidia's Blackwell chips. Mistral does not list a memory figure, but the hardware bill alone puts self hosting out of reach for most small teams.

Running a model locally also means owning its security. Our report on a Qwen model that tried to reach Alibaba Cloud during a test shows why teams should watch outbound network traffic from any open model they host.

Option two: pay per token through the API

What you get

For most developers, the hosted route is simpler. Mistral sells access through its API as mistral-large-4 and through Mistral Studio. The model is also on OpenRouter, on Nvidia hosted endpoints and in an Nvidia NIM container. Mistral named Microsoft as the first cloud provider to offer Large 4, through Microsoft Foundry.

What it costs

Mistral lists the API price at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Combined with the claim that Large 4 uses fewer tokens per task than GLM-5.3, that makes it one of the cheaper ways to run long coding agents. Token efficiency is worth watching closely as tools like GitHub Copilot move to credit based pricing.

How Mistral Large 4 handles security work

Mistral is pitching the model at defensive cybersecurity as much as coding. It reports a 51% score on the Artificial Analysis Cyber Index and an average refusal rate of 96.6% on malicious cyber requests.

The company is also candid about the risk. During the preview, Mistral wrote, "we see a potential for uplift to conduct real attacks," and said it was adjusting its guardrails. A specialized cyber version went only to Orange Cyberdefense, under vetting and contractual safeguards. Anyone hosting the open weights will be responsible for their own guardrails, which is a real difference from the managed API.

Which path fits which team

• Regulated companies and public agencies that cannot send data outside their walls: self hosting is the reason this release matters. Budget for an 8 GPU node and a security review.
• Startups and solo developers building coding agents: the API is the sensible start. The per token price is low, and there is no hardware to manage.
• Teams already on Azure: Microsoft Foundry may be the easiest way in, since billing and access stay in one place.
• Security teams: the benchmarks look promising, but wait for the technical report before trusting vendor numbers on attack resistance.

Mistral chief executive Arthur Mensch said the company has raised roughly $6 billion in three years, SiliconANGLE reported. Releasing its biggest model for free download is a bet that control, not just raw scores, is what buyers in Europe and beyond will pay for. For more model releases and AI policy news, visit our AI section.