What Mistral Large 4's open weights mean for coders

Mistral has released the downloadable weights of Mistral Large 4, its 1-trillion-parameter flagship model, nicknamed Le Chonk. The company announced it in We’re releasing the weights of Mistral Large 4. The model has been in public preview since earlier this month. Mistral says three more weeks of reinforcement learning moved it to the front of several open-weight coding benchmarks. The figure that matters most for you is about efficiency. According to Mistral, the model scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with the leading open model GLM-5.3, while using 47% fewer tokens per task on average.

The background you need

Weights and open weights. A trained language model is mostly a very large file of numbers called weights. Those numbers encode what the model learned. When a company releases the weights, you can download the model and run it on your own hardware. You can also fine-tune it on your own data, and nothing you send it has to leave your servers. Open weights is not the same as fully open source. You usually do not get the training data or the full training recipe.

An analogy: open weights are like getting a compiled program instead of its source code. You can run it, inspect how it behaves, and deploy it wherever you like. The analogy breaks in one place. You can keep training the weights on new data and change the model in ways that are hard to do with a compiled binary.

Parameters. The 1 trillion figure is the number of weights. It gives a rough sense of the model’s size and how much memory it needs. A model this size does not run on a laptop or a single consumer graphics card. It needs data-center hardware, usually several high-end GPUs working together.

Tokens. Models read and write text in chunks called tokens, which are roughly word fragments. Hosted APIs charge you per token. On your own hardware, tokens cost GPU time and electricity. Reasoning models also produce intermediate thinking tokens before they give an answer, and you pay for those too.

Coding agents. An agent is a model running in a loop. It reads files, runs commands, checks the output, edits code, and repeats until the task is done. Every step uses tokens, so one task can add up to a lot. Mistral’s own chart measures average tokens per task in millions.

Reinforcement learning (RL). After a model learns from examples (supervised fine-tuning, or SFT), RL trains it by rewarding good outcomes, such as code that passes its tests. Mistral credits its post-preview gains to scaling RL.

Sovereign AI. This is Mistral’s term for models an organization can host and control itself. The goal is often to keep data inside a particular jurisdiction. Mistral stresses that Large 4 was trained from scratch on European infrastructure.

What actually changed

Some things were already true before this release. Large 4 was already a 1-trillion-parameter, natively multimodal model. It was already available in public preview. What is new is that the weights are downloadable, and the released version is an improved model, not the preview.

According to Mistral, the changes since the preview include:

• Terminal-Bench 4.0 (coding tasks done through a command line): up 13 percentage points to 41%. Mistral says this leads open-weight models, and the model also produces 30% fewer output tokens.
• FrontierCode 2.0, a private benchmark from Cognition that checks whether a repository maintainer would merge the agent’s change: Large 4 leads with 19.7%.
• Three reasoning modes, so you can trade thoroughness against cost. Mistral describes the highest mode as meant for the hardest use cases that need high precision.
• Security behavior: an average refusal rate of 96.6% on malicious cyber requests. Mistral calls this the highest it measures among open models. On the B3 Agent Security Benchmark, the model resisted 97.1% of attacks, up from 93.3% in the preview.

Mistral also reports 51% on the Artificial Analysis Cyber Index, 73.3% on the Finch finance benchmark, and 84.5% on BrowseComp. In blind pairwise tests run by Surge Code, graders preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Among the models evaluated, it ranked behind only Claude Opus 5.5.

One item is not part of the public release. A specialized variant tuned for penetration testing and red-teaming is going only to vetted enterprise partners such as Orange Cyberdefense, under contractual safeguards.

What it means for curious technical readers

The Coding Agent Index result matters more than any single leaderboard lead. A 48% score tied with GLM-5.3 says Large 4 is about as capable. Using 47% fewer tokens per task says it gets there with roughly half the output. If your costs scale with tokens, whether through API billing or GPU hours, that difference shows up on every task. Benchmark leads tend to shrink as new models arrive. Spending half the tokens on the same work compounds across thousands of agent runs.

Think of two contractors who finish the same jobs equally well, but one bills about half the hours. This analogy also has a limit. If you self-host, much of your cost is the hardware you reserve, whether or not it is busy. Fewer tokens then means faster runs and more tasks per machine, not an automatically smaller bill. Per-token prices from hosting providers will also differ, so a model that uses fewer tokens is not guaranteed to be cheaper per task.

The Terminal-Bench result points the same way: a higher score with 30% fewer output tokens. Shorter outputs also tend to mean shorter waits in an agent loop.

What to watch next:

• Whether Artificial Analysis and others reproduce these figures on their own runs.
• Per-token prices from hosting providers that serve the open weights.
• How much the lower reasoning modes give up compared with the higher ones on your own tasks.

The security side is more complicated. Mistral says it watched cybersecurity use during the preview. It saw the model refusing black-box attacks on live targets, yet it writes that “we see a potential for uplift to conduct real attacks.” Mistral says it is changing its guardrails and will keep updating them in the coming months.

What we don’t know yet

• License terms. The announcement calls the model fully open-weight but does not spell out the license. That determines what you can do commercially.
• Hardware requirements. Mistral gives no guidance on the memory or GPUs needed to self-host a 1-trillion-parameter model.
• How guardrail updates reach downloaded copies. Mistral says it will keep updating its guardrails. The announcement does not say whether or how those updates apply to weights that people have already downloaded.
• The reasoning modes in detail. The source describes the top mode. It does not clearly lay out the cost and quality trade-offs of the other two.
• Independent replication. None of the results have been confirmed outside Mistral’s post yet.