Quick Answer: Mistral has released the open weights for Mistral Large 4, a 1-trillion-parameter multimodal model. Mistral reports it leads open-weight models on coding benchmarks and scores 51% on the Artificial Analysis Cyber Index. It offers three reasoning modes and can be self-hosted, which matters most for teams with strict data-residency rules.
The Mistral Large 4 open weights are now public, and that changes who can run this model. Open-weight models usually trail the closed frontier on the tasks developers care about most. Mistral argues that gap is closing for agentic coding and defensive security. As Mistral's announcement explains, the company trained the model from scratch on European infrastructure and published the weights after a public preview. The numbers are impressive, but they need context. Here's what they show, what they don't, and whether self-hosting a trillion-parameter model makes sense for your team.
What Mistral Large 4 Actually Is
Mistral Large 4, nicknamed "Le Chonk," is a natively multimodal model with roughly 1 trillion parameters. The public preview came first. Since then, Mistral says it ran about three more weeks of reinforcement learning, focused on coding and defensive cybersecurity. The weights being released today belong to that improved version, not the preview.
That distinction matters more than it looks. If your team benchmarked the preview in the spring or summer, those results don't describe what you'll download now. Re-run your own evaluations on the released checkpoint before you make a decision based on earlier numbers.
Mistral also claims substantial gains in reasoning efficiency, meaning more performance per token and lower inference cost. Hold that claim loosely for now. Efficiency is the number that determines your bill, so it deserves a test on your own workload, not just a glance at a chart.
Coding Results: Accuracy Is Only Half the Story
The most common advice when comparing coding models is to pick the highest score. In agent loops, that advice is incomplete. An agent that solves a task with half the tokens runs roughly twice as many tasks on the same budget, and the savings compound across thousands of runs.
That's why Mistral's token-efficiency claim matters. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, effectively tied with GLM-5.3, the leading open-weight model. It uses 47% fewer tokens per task on average. On Terminal-Bench 4.0, it reached 41%, up 13 percentage points from the preview's 28%, while reducing output tokens by 30%.
Cognition's FrontierCode 2.0 is a different kind of test. It judges end-to-end code quality by asking whether a repository maintainer would merge the change. Mistral Large 4 leads that benchmark with 19.7%. The number looks modest, but the benchmark is hard, and it's closer to real pull-request review than most leaderboards. Note that FrontierCode 2.0 is Cognition's private benchmark, so you can't independently reproduce it yet.
Here's where most teams stop reading, and they shouldn't. Human judgment matters too. In Surge Code's blind pairwise evaluations, graders preferred Mistral Large 4's answers over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher among the models tested.
Three Reasoning Modes and When to Use Each
Mistral Large 4 now exposes three reasoning levels. Each one trades cost against precision:
- Low: simple tasks. This replaces the old "None" mode.
- Medium: everyday coding tasks.
- High (default): the hardest problems that need precision and thorough reasoning.
The default is the detail to watch. High is the setting most people will run without thinking about it, and it's the most expensive. The lazy assumption is that more reasoning always produces better code. In practice, a routine rename or a docstring update doesn't need high reasoning, and paying for it adds cost without improving the result.
A workable pattern is to run routine edits on Low or Medium and escalate to High only when tests fail or the change touches concurrency, auth, or data migrations. That's a trade-off you can measure. Track pass rate and tokens per task for each mode across a sample of your real tickets. That data will tell you more than any vendor chart.
Defensive Security: Capability Without Blanket Refusals
Security is where the release gets more interesting, and more delicate. Mistral reports 51% on the Artificial Analysis Cyber Index, which covers finding vulnerabilities in a codebase, reproducing them, and patching them without breaking existing functionality. The hard part isn't finding bugs. It's doing all three steps without breaking the code.
The tension is between usefulness and safety. Many models refuse anything that smells like security work, which makes them useless to defenders. Mistral claims the opposite problem: its guardrails separate defensive tasks from offensive ones. The company reports an average refusal rate of 96.6% on malicious cyber requests, which it describes as the highest it measures among open models. Those requests draw on JailbreakBench, StrongREJECT, and AgentHarm prompts.
Treat both numbers as vendor measurements. Mistral ran them, and the refusal benchmarks are its own selection. Still, outside partners have reported results. Booz Allen said the model matched vulnerability-detection recall against other frontier cyber models on its internal source-code analysis benchmark. ENISA says it is testing Mistral Large 4 to find vulnerabilities, audit code, and analyze defensive logs.
Cisco's Amy Chang made a related point in the announcement: as adversaries use AI to speed up their operations, defenders need tools they can inspect and control. That argument is the bridge to the next section.
Benchmarks at a Glance
The table below puts the headline results side by side. Each row lists what the metric actually tests, since the labels hide real differences.
| Benchmark | Mistral Large 4 result | What it measures | Caveat |
|---|---|---|---|
| FrontierCode 2.0 | 19.7% | Whether a maintainer would merge the change | Cognition's private benchmark |
| Artificial Analysis Coding Agent Index | 48% | Agentic coding quality, tied with GLM-5.3 | Uses 47% fewer tokens per task |
| Terminal-Bench 4.0 | 41% | Terminal-based coding workflows | Up from 28% in preview; 30% fewer output tokens |
| Artificial Analysis Cyber Index | 51% | Find, reproduce, and patch vulnerabilities | Requires functional patches, not just findings |
| Malicious cyber refusal rate | 96.6% | Refusing offensive cyber requests | Mistral's own measurement |
Notice that the strongest claims are relative (leads, ties, fewer tokens), not absolute. A 19.7% score on a hard merge-readiness benchmark is a good result, not a solved problem.
Self-Hosting and Sovereignty: The Cost Question Most Teams Skip
Open weights sound like a cost win, and that's the lazy read. A 1-trillion-parameter model needs serious GPU capacity to serve. For low volume, a hosted API is almost always cheaper once you count hardware, utilization, and the engineering time to keep inference healthy. Self-hosting pays off when you have steady, high-volume traffic or when data can't leave your environment.
That second reason is where the sovereignty argument carries weight. Rubrik's Anneka Gupta described European governments and regulated industries needing options that don't route their data to other countries. The Mistral Large 4 open weights make that option real: you can host the model in your own environment and keep data under your control.
If you're evaluating self-hosting, plan for these items:
- Hardware: size your GPU fleet against your peak concurrency, not your average load.
- Serving stack: frameworks like vLLM and NVIDIA's TensorRT-LLM are common choices for large-model inference. Check the release notes for supported configurations.
- Quantization trade-offs: lower-precision weights cut memory use but can shift quality, so re-run your coding benchmarks after any change.
- Licensing: read the license file that ships with the weights before you put anything into production.
- Monitoring: log token counts, latency, and refusal rates so you can catch regressions after updates.
So how to self-host Mistral Large 4 for coding agents? Start small. Run a pilot on one internal repository, measure tokens per task and pass rate, then decide whether the infrastructure bill beats your current API spend.
What the Benchmarks Don't Tell You
Every claim in this release comes with a source, and most of those sources are the vendor or a partner. FrontierCode 2.0 is private. The Surge Code preference results are blind, which helps, but they still reflect one evaluator's prompts. The Terminal-Bench figure is labeled as self-reported in the release.
That doesn't make the results wrong. It does mean you should check them against your own codebase. The metric that matters most is how often the model produces a change your team would merge without rework. Benchmark rankings can't tell you that.
There's also a fair question about open release and misuse. Mistral says its guardrails handle the split between defensive and offensive work, and it published refusal data to back that up. Researchers can test that claim directly now that the weights are public. Whether that transparency is enough for your risk tolerance is a decision for your security team, not a leaderboard.
Frequently Asked Questions
What are the Mistral Large 4 open weights?
The Mistral Large 4 open weights are the downloadable model parameters for Mistral's 1-trillion-parameter multimodal model, released after a public preview. They let organizations host, customize, and inspect the model on their own infrastructure. Mistral says the released version includes about three weeks of additional reinforcement learning on coding and cybersecurity.
How do I choose between the Low, Medium, and High reasoning modes?
Use Low for simple tasks, Medium for everyday coding, and High (the default) for the hardest problems that need precision. Higher reasoning costs more tokens, so High on routine edits wastes budget. Measure pass rate and tokens per task on your own tickets to set the right default for your team.
why do open-weight models matter for sovereign AI?
Open-weight models let organizations run AI under their own control, so sensitive data never has to leave their environment. This matters for regulated industries and governments with strict data-residency requirements. It also lets security teams audit how a model behaves rather than trusting a closed API.
Is self-hosting Mistral Large 4 cheaper than using an API?
Not always. A model of this size needs significant GPU capacity, and low-volume workloads often cost less through a hosted API. Self-hosting usually makes sense only with steady, high traffic or strict data-control requirements. Model the total cost, including hardware, utilization, and engineering time, before deciding.
Pick one real coding task from your backlog this week and run it through Mistral Large 4 on Low, Medium, and High, then record the tokens and whether the change would pass review. Pass this to whoever on your team is wrestling with an agent bill that keeps climbing. Your own numbers will settle the open-weight question faster than any leaderboard.