Mistral Large 4 has moved beyond the promise of a future weight release. On October 29, 2026, Mistral made an improved version of its flagship model, also known as Le Chonk, available to download.
For developers and businesses, the practical change is the ability to choose how the model is operated and adapted. The release also brings updated API pricing and a licence that sets specific conditions for model-service businesses.
DemystIA covered the initial announcement in our Mistral Large 4 public-preview article. This follow-up examines what the downloadable release makes possible.
What do downloadable weights actually give you?
Weights are the numerical parameters learned during training. Access to them allows an organisation to run a model using compatible software and suitable infrastructure, and to adapt it within the terms of its licence.
Mistral’s official model card describes a roughly one-trillion-parameter Mixture-of-Experts model. It accepts text and images and produces text, with support for tool calls and structured outputs.
For a software team, that creates room to build an assistant around its repositories, documentation and development tools. The team can decide which systems the assistant may access and where the model runs.
Open weights provide a deployment option; a complete working agent still needs software, permissions and an evaluation process. Downloading the model also does not, by itself, provide the full training data or the complete process used to create it.
The coding claim: similar scores, fewer tokens
In its release announcement, Mistral reports a 48% score on the Artificial Analysis Coding Agent Index, approximately matching GLM-5.3 while using 47% fewer tokens per task on average.
That is a comparison reported by the model’s developer. DemystIA has not reproduced the evaluation. Its scope also matters: fewer tokens on that test do not automatically translate into the same saving on another codebase or hosting service.
The model card lists three reasoning settings: low, medium and high, with high as the default. A useful trial would compare them on the same tasks, recording successful changes, response time and total cost.
A model that finishes a task with fewer tokens could make repeated development work cheaper. The saving needs to survive code review: retries and corrections can consume the apparent advantage.
Can you run the full model on your own hardware?
Hugging Face’s release guide lists official FP8 and NVFP4 variants, alongside community conversions. Those formats affect how the numerical weights are represented and how much memory they occupy.
The scale remains substantial. As an order-of-magnitude calculation, one trillion parameters stored at eight bits each represent about one terabyte of raw weights. At four bits, the corresponding figure is roughly 500 gigabytes.
These are illustrative calculations for weights alone, excluding metadata and runtime memory. They are not deployment specifications. Actual requirements depend on the checkpoint, serving software, context length and number of simultaneous users.
A Mixture-of-Experts architecture activates selected parts of a model for each token. That can reduce computation, while the full set of weights still needs to remain accessible to the system.
For a small team, testing a hosted endpoint first can establish whether the results justify a dedicated infrastructure project.
What could it change for defensive security?
Mistral positions the release for vulnerability reproduction and patching, alongside coding. That makes it relevant to teams maintaining applications and investigating suspected defects.
Consider a hypothetical team maintaining a WordPress plugin. A contained evaluation could give the assistant a copy of the repository, a documented permission-checking defect and a disposable test environment. The task would be to explain the faulty path, propose a focused patch and add a regression test.
A reviewer would check whether the test exposes the original defect, whether the patch resolves it and whether legitimate users retain the expected functionality. These checks establish more than a convincing explanation alone.
This is a suggested evaluation scenario, not a result observed with Le Chonk. Our article on AI-assisted PHP vulnerability research provides further context on the relationship between finding a flaw and verifying a correction.
What does the API cost?
Mistral’s model documentation, checked on November 1, 2026, lists the following USD prices:
| Token category | Price per million tokens |
|---|---|
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |
Using these rates, a hypothetical workload billed for 100,000 uncached input tokens and 10,000 output tokens would cost $0.091 in model-token charges: $0.07 for input plus $0.021 for output.
The example excludes additional tools, retries, taxes and review time. It also assumes those figures cover all billed input and output tokens, including any billed reasoning tokens.
For an agent working through several steps, measure the complete run. A single final answer may follow multiple model requests, tool results and unsuccessful attempts.
Read the licence before building a hosted service
The Mistral Ecosystem Open-Weight License permits use, modification and redistribution subject to its conditions, including attribution.
It requires a separate commercial agreement when the licensee or an affiliate operates a Model-as-a-Service business and uses the model commercially, with specified exceptions including internal use and access through Mistral’s official products. Its definition also excludes certain products that embed model capabilities within specific features.
The practical implication is that an internal assistant and a service giving customers broad control over model inference can fall under different terms. A provider planning commercial hosting should check the published definition against its intended product.
How to decide whether Le Chonk is useful for your team
DemystIA recommends starting with a small collection of completed tasks from the team’s own backlog. Preserve the original problem, the expected behaviour and the checks used to accept the human-written solution.
Run the same tasks with the existing assistant and the new model. Record:
- Accepted results: changes that satisfy the requirements and pass review.
- Correction effort: time spent explaining mistakes or rewriting the proposed solution.
- Full-run cost: model usage, tools and any repeated attempts.
- Operating effort: integration, monitoring and infrastructure maintenance.
Include an unfamiliar repository and an incomplete bug report. These reveal how the assistant handles missing information, which a tidy demonstration may conceal.
Our assessment is that this release makes Mistral’s deployment proposition more concrete. Teams can now inspect the actual distribution and licence while testing the hosted model. The adoption decision should follow from accepted work and operating requirements measured together.
Sources
- Mistral AI — Le Chonk weight-release announcement, October 29, 2026.
- Mistral AI — Official model card, capabilities and reasoning settings.
- Hugging Face — Welcome Le Chonk: Mistral Large 4, October 29, 2026.
- Mistral documentation — Model specifications and API pricing.
- Mistral AI — Mistral Ecosystem Open-Weight License.
Sources checked on November 1, 2026. This article analyses published documentation. Calculations and evaluation scenarios are illustrative; DemystIA has not performed a hands-on benchmark of the model.