# Mistral Opened a 1T Model. The Catch Is the MaaS Licence

> Mistral released the weights of its 1 trillion parameter Large 4 under a modified MIT licence, with one clause that decides who can sell access to it.

- Canonical: https://zeroraised.com/blog/mistral-large-4-open-weights
- Published: 2026-10-30
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, developer-tools, comparison, bootstrapping

Mistral has published the weights of Mistral Large 4, a 1 trillion parameter natively multimodal model it nicknames Le Chonk, after a public preview earlier in October ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). The model is strong, especially for coding, and the licence is genuinely permissive. The catch is one sentence: operating a model-as-a-service business on top of it requires a separate commercial licence with Mistral ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). That line splits founders into two groups.

**TL;DR**

- Mistral Large 4 is open weight at about 1T total parameters and 49B active per token, with a 1M token context ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)).
- Mistral calls it the best open-weight model for coding and notes it leads FrontierCode 2.0 at 19.7% ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)).
- The Mistral Ecosystem Open-Weight License is a modified MIT licence: free to use, modify, fine-tune and commercialise.
- Selling access to the model as a service needs a separate commercial licence ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)).
- API pricing is $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07 ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)).

**Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 30 Oct 2026; ZeroRaised has no commercial relationship with the companies named.

**A service-restricted open-weight licence** is a permissive model licence that grants free use, modification, fine-tuning and commercialisation, but requires a separate paid agreement when a company sells access to the model itself.

## How strong is Mistral Large 4 for coding?

Mistral calls it the best open-weight model for coding, and its evidence starts with FrontierCode 2.0, a private benchmark from Cognition, where the model leads at 19.7% ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). Private benchmarks deserve caution, but the rest of the stack is consistent. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). On Terminal-Bench 4.0 it reaches 41%, which is 13 points higher than the preview, with 30% fewer output tokens ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)).

In Surge's blind pairwise coding comparison it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and behind only Claude Opus 5.5 ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). That is a credible position for an open-weight model, and blind preference tests measure what buyers actually feel.

The model has three reasoning modes, Low, Medium and High, with High as the default ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). On security and agent tests it reports a 51% score on the Artificial Analysis Cyber Index, a 96.6% average refusal rate on malicious cyber requests, 84.5% on BrowseComp and 97.1% on B3 Agent Security, up from 93.3% in the preview ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)).

## What does the licence actually allow?

The Mistral Ecosystem Open-Weight License is a modified MIT licence. It is free to use, modify, fine-tune and commercialise the model ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). That covers the ordinary startup pattern: build a product, call the model, wrap it in a workflow, charge for the outcome.

The single condition is aimed at a narrower business: operating a model-as-a-service business requires a separate commercial licence with Mistral ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). The distinction sounds small until you map it onto your product. A coding assistant that helps a customer is one thing. A bare endpoint that sells tokens on top of the model is another.

The boundary is a legal question, not a blog post. If your business is close to it, read the licence text and get advice. The [open-weight versus closed model comparison](https://zeroraised.com/blog/open-weight-vs-closed-ai-models) covers the capability side of the same decision.

## What does it cost to run?

Mistral sells API access at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). The model is available through Mistral Studio, the Mistral API, OpenRouter, Hugging Face with vLLM and SGLang, Microsoft Foundry and NVIDIA NIM ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). That spread of providers is the useful part: you can switch hosts without changing the model.

Self-hosting is a different calculation. The model has about 1T total parameters and 49B active per token, 128 routed experts with 4 active per token plus 1 shared expert, a 1M token context, text and image input, text output and support for more than 160 languages ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)). It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)).

The serving notes explain why this is not a laptop model. The recommended example uses tensor parallel size 8, a node of eight GPUs, and a quantized NVFP4 checkpoint targets Blackwell GPUs ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)). The SGLang cookbook puts FP8 on 8x B300 and NVFP4 on 4x GB300 ([SGLang](https://docs.sglang.io/cookbook/autoregressive/Mistral/Mistral-Large-4)). A DSpark speculative decoding draft head is included, and Mistral measured up to 1.9x higher throughput with it ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)).

Those are data-centre bills. On the Artificial Analysis Intelligence Index v5 it scores 26, the highest among western open-weight models, though leading Chinese open-weight models score higher on that aggregate ([Hugging Face](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)). The realistic choice for a small team is to call it through an API, or run a smaller model locally and route only the hard cases upward. The [LLM cost guide](https://zeroraised.com/blog/how-to-reduce-llm-costs) and the [guide to running a model locally](https://zeroraised.com/blog/how-to-run-an-llm-locally) cover both paths.

## Should a bootstrapped founder build on it?

If your product is a coding agent or a multi-file engineering assistant, this is now a serious option. The token efficiency is the story: matching a rival while using 47% fewer tokens per task changes the cost per completed job, not just the cost per token ([Mistral](https://mistral.ai/news/mistral-large-4-oss/)). Price per million tokens misleads when one model needs half the tokens. Measure cost per task on your own workload.

If your product is a thin wrapper that resells inference, the licence is the decision. Free weights do not mean free rein, and the [open-weight licence guide](https://zeroraised.com/blog/open-weight-model-licenses) walks through the alternatives. If you are still choosing a provider, the [LLM API guide](https://zeroraised.com/blog/how-to-choose-an-llm-api) covers the questions that matter before you write the integration.

## What this means for bootstrapped founders

- **Separate the model from the business model.** Using an open-weight model inside a product is permitted. Selling the model itself as a service is the line that triggers a commercial agreement.
- **Compare cost per task, not per token.** A model that finishes the job in fewer tokens can be cheaper even at a higher sticker price.
- **Do not plan to self-host a 1T model.** Eight-GPU nodes are a capital commitment. Use an API or a hosted provider and keep a smaller local model for cheap, bounded work.
- **Keep a model-agnostic layer.** With availability across OpenRouter, Hugging Face, Foundry and NIM, switching providers should be a configuration change, not a rewrite.
- **Read the licence before the launch, not after.** A licence clause discovered after shipping is an expensive surprise.

## Where ZeroRaised fits

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have built something on Mistral Large 4, you can [submit your product](https://zeroraised.com/submit). The [pricing page](https://zeroraised.com/pricing) explains the current options, and there are no promises of traffic or results.

If you want more analysis like this, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter). It is a plain email, not a growth hack.

## Sources

- [Mistral: Mistral Large 4 open weights](https://mistral.ai/news/mistral-large-4-oss/)
- [Hugging Face: Mistral-Large-4-1T-A52B model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)
- [SGLang cookbook: Mistral Large 4](https://docs.sglang.io/cookbook/autoregressive/Mistral/Mistral-Large-4)

## Frequently asked questions

### Is Mistral Large 4 really open source?

Mistral publishes the model weights under the Mistral Ecosystem Open-Weight License, a modified MIT licence. It is free to use, modify, fine-tune and commercialise, but the licence adds one condition for companies that sell access to the model itself.

### What is the model-as-a-service clause?

Operating a model-as-a-service business on top of Mistral Large 4 requires a separate commercial licence with Mistral. Building a product that uses the model is not the same as reselling inference on the model, and founders should read the licence text before shipping.

### Can I fine-tune Mistral Large 4 and keep the result private?

Mistral says the licence lets you use, modify, fine-tune and commercialise the model. Check the current licence text for attribution, notice and model-as-a-service terms before you ship a fine-tune.

### How much does the API cost?

Mistral lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Self-hosting is a separate calculation because the model needs a node of eight GPUs in the recommended example.

### Can a solo founder self-host it?

Technically the model card recommends tensor parallel size eight, and the SGLang cookbook shows FP8 running on eight B300 GPUs. That is a capital cost most bootstrapped teams will not pay, so the API or a hosted provider is the realistic route.