Mistral Large 4 arrived on October 6, 2026, and it immediately changed the conversation about open models. For example, Mistral calls it the largest and most capable model in its history, and the scale supports that claim. However, scale alone never tells the full story, because developers also need price, speed, and a clear path to deployment. Therefore, this guide cuts through the launch noise and explains what the model does, what it costs, and where it fits.
In addition, Mistral released the model first as a public preview through its API, and it promised the open weights by the end of October 2026. As a result, early adopters can test the preview today, while teams that need self hosting should wait for the weight release. Meanwhile, buyers should compare real workloads instead of headline parameter counts.
What Is Mistral Large 4?
Mistral Large 4 is a natively multimodal mixture of experts model from Mistral AI, and Mistral nicknamed it Le Chonk. For example, it reads text and images from the start, rather than bolting vision on as a separate tool. In addition, it supports more than 160 languages, including every official language of the European Union. Therefore, European teams gain a strong local option for multilingual support, coding, and document work.
Mistral trained the model from scratch on about 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. However, training scale does not guarantee easy deployment, because the full model still needs serious infrastructure. As a result, most small teams should start with the API and treat self hosting as a later step.
Mistral describes the model as its largest and most capable model to date in the official announcement, and that framing matches the jump from Mistral Large 3. For example, Large 3 offered 675 billion total parameters and 41 billion active parameters, while the new model reaches about 1.05 trillion total parameters. In addition, the context window grows from 256,000 tokens to 1 million tokens. Therefore, long documents, large codebases, and extended agent sessions become far more practical.
How the Mixture of Experts Design Works
A mixture of experts model splits its network into many specialist subnetworks, and a router activates only a few experts for each token. For example, one expert may handle code syntax, while another handles legal language or image details. However, the router must choose well, because a poor choice can waste the benefit of scale. Therefore, Mistral tuned the routing for coding, agentic work, and multimodal understanding.
In addition, the design explains the price. Mistral lists about 52 billion active parameters per token in its announcement, while its model card reports 49 billion active parameters plus embeddings. As a result, you should treat the active count as roughly 49 to 52 billion, not as a single exact figure. Meanwhile, the 1.6 billion parameter vision encoder handles image input, so screenshots, diagrams, and scanned documents can enter the same workflow as text.
Key Features and Benchmarks
The model combines a huge context window, native vision, and strong enterprise skills. For example, it supports coding, agentic workflows, finance, law, and cybersecurity in one model. However, vendor benchmarks always need a second look, because companies choose tests that flatter their strengths. Therefore, treat the numbers below as a starting point, then run your own tests on your own data.
Coding and Agent Performance
Mistral reports a Coding Agent Index score of 49.8 percent, and it places that result ahead of DeepSeek V4 Pro and Qwen3.8 Max. For example, developers can use the model for code review, debugging, and multi step refactoring. In addition, a Surge AI human evaluation scored the preview at 3.74 out of 5 for coding quality, behind Claude Opus 5 at 4.22. Therefore, the model looks strong for coding, but it does not yet lead the field on human judged quality.
Cybersecurity Strength
Cybersecurity stands out as the clearest strength. For example, Mistral reports 93 percent on Cybench and 82 percent on CyberGym-E2E, and it says several closed frontier models score near zero because they refuse those tasks. However, refusal behaviour differs by policy, so a high score does not prove the model suits every security workflow. As a result, security teams should test detection rules, vulnerability triage, and malware analysis in a controlled lab first. Meanwhile, Mistral says it red teamed the model with cybersecurity leaders and vetted partners before the weight release.
Mistral Large 4 Pricing Explained
The pricing starts at 1.36 dollars per million input tokens and 4.18 dollars per million output tokens, with cached input at 0.14 dollars per million tokens. For example, a team that reuses long prompts can cut input costs sharply with caching. However, output tokens still dominate the bill for long reports and code generation, because output costs about three times the input rate. Therefore, teams should set output limits, use concise prompts, and cache stable context wherever possible.
In addition, Mistral showed a preview discount at half those rates, at 0.68 dollars for input, 0.07 dollars for cached input, and 2.09 dollars for output. As a result, early testing costs less than the standard list price, but teams should budget at the standard rates for production. Meanwhile, batch pricing can reduce costs further for work that does not need an instant reply. Finally, always confirm the current rate in Mistral Studio before you sign off a budget, because preview pricing can change.
Comparison Table: Rivals at a Glance
| Model | Total parameters | Context | Input price per million tokens | Best for |
|---|---|---|---|---|
| Mistral Large 4 | 1.05 trillion | 1 million tokens | 1.36 dollars | Open enterprise agents and security |
| Mistral Large 3 | 675 billion | 256,000 tokens | Lower | Cheaper legacy Mistral workloads |
| NVIDIA Nemotron 3.5 Super VL | 120 billion | 1 million tokens | 0.15 dollars | Multimodal agents on NVIDIA hardware |
| Claude Opus 5 | Not disclosed | Large | Higher | Human judged coding quality |
For another open model built for agents, see our guide to NVIDIA Nemotron 3.5, because it shows how a smaller multimodal model can still serve enterprise agents. In addition, if speed matters more than scale, our Mercury 3 coverage shows how a diffusion model changes latency for real time work. Therefore, choose Mistral Large 4 for scale and openness, and choose a rival when price or raw speed matters more.
Strengths and Limitations
The strengths start with scale, openness, and European hosting. For example, teams gain a 1 million token context, native vision, and broad language support in one API. In addition, cybersecurity performance gives security teams a practical reason to test the model now. Therefore, regulated European organisations gain a credible alternative to closed United States models.
However, the limitations matter just as much. First, the open weights had not shipped at launch, so teams could not self host on day one. Second, Mistral had not confirmed the final licence at launch, and the licence decides commercial use. Third, the full model will need heavy infrastructure, so self hosting will suit large teams first. As a result, small teams should treat the API as the main path, and large teams should plan hardware, compliance, and support before they commit.
Who Should Use It
Enterprise developers should test Mistral Large 4 first when they build coding agents, document agents, or security tools. For example, a bank can triage long reports, a law firm can review contracts, and a security team can prioritise vulnerabilities. In addition, European companies that need data residency and multilingual support gain a clear advantage. Therefore, the model suits production teams with real compliance needs, not just hobby projects.
Meanwhile, startups and solo builders should test it with a narrow pilot. For example, pick one workflow, measure accuracy and cost for two weeks, and compare the result with a cheaper small model. As a result, a pilot protects your budget and gives you evidence for a larger rollout.
Frequently Asked Questions
What is Mistral Large 4 pricing?
Mistral Large 4 pricing lists at 1.36 dollars per million input tokens, 0.14 dollars for cached input, and 4.18 dollars per million output tokens. For example, caching long prompts can cut input costs sharply. However, always check Mistral Studio for the current rate, because preview discounts can end.
How do Mistral Large 4 benchmarks compare with rivals?
Mistral Large 4 benchmarks look strongest in cybersecurity, with 93 percent on Cybench, and solid in coding, with a 49.8 percent Coding Agent Index score. For example, human evaluators still ranked Claude Opus 5 higher for coding quality. Therefore, run your own benchmark on your own tasks before you choose.
How does Mistral Large 4 compare with Large 3?
Mistral Large 4 vs Large 3 shows a clear jump in scale, context, and vision. For example, total parameters rise from 675 billion to about 1.05 trillion, and context grows from 256,000 tokens to 1 million tokens. In addition, the new vision encoder strengthens image understanding for enterprise documents.
When will Mistral Large 4 open weights arrive?
Mistral Large 4 open weights were promised by the end of October 2026. For example, Mistral said it would release a final version with the weights after red teaming. However, teams should confirm the licence terms at release, because the licence controls commercial self hosting.
What is the Mistral Large 4 context window?
The Mistral Large 4 context window reaches 1 million tokens in Mistral documentation. For example, that window can hold large codebases, long contracts, and extended agent histories. However, some providers capped the preview lower, so confirm the limit in your chosen API before you send very long prompts.
What does Le Chonk mean?
Le Chonk is the playful internal nickname Mistral used openly in its launch material. For example, the name signals a large, capable model with personality. However, the nickname changes nothing about pricing or performance, so judge the model on tests, not on branding.
Final Verdict
This model earns attention for a simple reason: it pairs trillion parameter scale with a practical API price and a clear European hosting story. For example, security teams, coding teams, and multilingual enterprises can test real value this month. However, the open weight promise still needs a licence and a final checkpoint, so buyers should treat self hosting as a plan, not a present fact. Therefore, start with a focused API pilot, measure cost and accuracy, and expand only when your own evidence supports the move. Finally, keep watching the weight release, because that release will decide whether this model becomes a true open alternative or remains a strong hosted option. Independent coverage also tracks the preview pricing and weight timing for teams that want a second source.
