Open Weights Are Not Open Business Models: Mistral Large 4 and the Licensing Trap

The Release

On October 29, 2026, Mistral published the full weights of Mistral Large 4 — nicknamed "Le Chonk" by the company itself — on Hugging Face, three weeks after the model first appeared as an API-only preview on October 6. It is a natively multimodal mixture-of-experts with about 1.05 trillion total parameters, roughly 49 billion active per token. It accepts text and images, answers in text, and carries a 1-million-token context window. Mistral calls it the top open-weight model for coding, reporting 41% on Terminal-Bench 4.0 — up from 28% in the preview — and 48% on the Artificial Analysis Coding Agent Index.

Mistral says it is the first trillion-parameter model ever released with open weights. The weights are on Hugging Face as Mistral-Large-4-1T-A52B, and hosted API pricing is $0.70 per million input tokens, $0.07 cached, and $2.10 output. But the most important part of this release is not the parameter count. It is the licence.

The Licence, Read Carefully

Large 4 ships under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. On its face, that sounds generous: the weights are free to use, modify, fine-tune, and commercialise. A founder reading the headline could reasonably conclude that they can build a business on this model without ever talking to Mistral.

That conclusion would be wrong. The licence contains one condition, and it is the condition that matters most for anyone building an AI company. If your business operates a Model-as-a-Service offering — if you sell access to the model itself as a hosted service — you must enter into a separate commercial licence with Mistral before using the model commercially. The licence sets no revenue threshold. There is no small-business carve-out, no free tier, no "under $1 million in revenue" exemption. If you host the model for third parties and charge for it, you need a deal.

The condition does not apply to internal use — any use that does not make the model or your versions of it available to third parties — or to use through Mistral's own products. A company that downloads Large 4, fine-tunes it, and runs it inside its own product is fine. A company that wraps the model in an API and sells tokens to other developers is not fine without a commercial agreement.

Why This Matters More Than the Benchmarks

This is the licensing trap, and it is becoming the defining commercial question of the open-weight era. "Open weights" and "open source" are not the same thing, and the gap between them is where business models live or die.

A truly open-source licence, like Apache 2.0, lets anyone use the software for any purpose, including selling it as a service. A modified MIT licence with a Model-as-a-Service condition is something else: it is open for users and closed for competitors. Mistral is happy for you to build a coding assistant, a legal research tool, or an internal automation system on Large 4. It is not happy for you to become a cheaper Mistral.

From Mistral's perspective, this is entirely rational. Training a trillion-parameter model costs an enormous amount of money, and if anyone could download the weights and immediately undercut Mistral's own hosted API, the company would have no way to recover its investment. The MaaS condition is how Mistral funds the next model while still giving the weights away. The trap for founders is not that the licence is unfair — it is that it is easy to miss.

The Practical Checklist for Founders

If you are building on open-weight models, Large 4 or any other, here is the discipline that keeps you out of the trap.

Read the licence before you build, not after you launch. The licence is the business model. It tells you what you can sell, what you can host, and what requires a negotiation. Budget an hour with an AI-licensing lawyer before you commit your architecture to a model.

Know which side of the line you are on. The decisive question is simple: are you selling the model, or are you selling something built with the model? An internal tool or a product with the model embedded is on the safe side. An API that sells tokens or a hosted inference service is on the side that needs a deal.

Price in the hardware. Even the open side is not free. The full FP8 weights are about 1.06 terabytes, and the NVFP4 version is around 610 gigabytes. Mistral's own serving examples run on eight GPUs. The download is free; the electricity and the GPUs are not.

Test benchmarks on your own workload. Mistral's numbers are useful signals, not guarantees, and the technical report is still to come. A model that is best for coding agents may not be best for your use case.

The Bottom Line

Mistral Large 4 is a genuine gift to developers: a trillion-parameter model, free to download, modify, and build products on. But it is a gift with a boundary, and the boundary is the business model. Open weights are not open business models. The founders who thrive in the open-weight era will be the ones who read the licence as carefully as they read the benchmark table — and who build on the side of the line that belongs to them.

Angeled.com connects founders with investors.