What is Mistral Large 4?
Mistral Large 4, also called ML4, is a general-purpose multimodal mixture-of-experts model released in public preview on October 6, 2026. Mistral documents 1.05 trillion total parameters, 52 billion active parameters and a 1.6-billion-parameter vision encoder.
The model accepts text and images and produces reasoning and text. Mistral positions it for coding, agentic workflows, cybersecurity, science, finance, law and visually grounded work. It is served through Mistral Studio and the Mistral API while the company prepares downloadable weights.
QUICK VERDICT: Mistral Large 4 is Mistral AI's largest and most capable model so far, with one-million-token context, native vision and strong agentic positioning. The API is usable now, but it remains a public preview and the promised open weights were not yet available when this guide was checked.
Field | Verified value |
|---|---|
Provider | Mistral AI |
Release date | October 6, 2026 |
Status | Public preview |
License | Hosted preview now; open weights announced as coming soon |
API model IDs | mistral-large-4 and mistral-large-4-0 |
Context window | 1 million tokens |
Architecture | Granular mixture of experts |
Parameters | 1.05T total; 52B active |
Vision encoder | 1.6B parameters |
Modalities | Text and image input; reasoning and text output |
Standard API price | $1.36 input, $0.14 cached input and $4.18 output per million tokens |
Last verified | October 22, 2026 |
What changed from Mistral Large 3
Mistral Large 4 expands substantially beyond Large 3 in scale, context and multimodal design. It moves from the previous generation's open release to a staged rollout: a hosted public preview first, followed by promised weights and additional architecture and post-training details.
Area | Mistral Large 4 change |
|---|---|
Scale | 1.05T total parameters with 52B active per token |
Context | One-million-token context window |
Vision | Native multimodality with a 1.6B vision encoder |
Agentic work | Broader emphasis on long-running coding and business workflows |
Deployment | European-hosted preview with planned open weights and self-deployment |
Lifecycle | Public preview, so specifications and behavior may still change |
A newer model is not automatically the right migration target. Compare Mistral Large 4 with Mistral Large 3 using the same prompts, tools, source material, effort controls and acceptance rubric. Record quality, latency, retries, review time and total cost per accepted result.
Capabilities and best use cases
Mistral documents structured outputs, function calling, document Q&A, prefix completion, Chat Completions, batching, agent conversations and connectors. The launch also emphasizes visual grounding, large engineering drawings, geospatial imagery and professional artifacts such as spreadsheets and documents.
Capability | Practical use |
|---|---|
Agentic coding | Repository understanding, terminal work, debugging and implementation |
Cybersecurity | Authorized vulnerability research, malware analysis and detection engineering |
Multimodal analysis | Documents, charts, engineering drawings, scenes and geospatial images |
Knowledge work | Finance, legal, document and spreadsheet workflows |
Science and math | Scientific coding, simulation, applied mathematics and technical analysis |
Sovereign deployment | Planned open weights for private-cloud and on-premise operation |
The strongest documented fit is multimodal enterprise agents, coding, cybersecurity, scientific work and European or self-hosted deployment planning. Reliability still depends on retrieval, tool permissions, prompt design, output validation and escalation policy. Test the complete system instead of evaluating the model name in isolation.
Benchmark evidence
Mistral reports results across coding, cyber, automation, professional work and safety. Examples include 61.7% on DeepSWE v1.1, 59.9% on AutomationBench and 93.3% attack resistance on Lakera's B3 benchmark. These are launch-page results with varying external and internal evaluators, so each should be interpreted with its stated harness and evidence boundary.
Evaluation area | Reported evidence | How to interpret it |
|---|---|---|
DeepSWE v1.1 | 61.7% | Mistral-reported agentic coding result |
SWE-Atlas-QnA | 59.4% | Repository-understanding evaluation |
Terminal-Bench 4 | 28.3% | Complex command-line tasks |
AutomationBench | 59.9% | Provider-reported result across 657 business workflows |
Cybench | 93% | Provider-reported result on 40 cybersecurity exercises |
Lakera B3 | 93.3% attacks resisted | Public indirect prompt-injection benchmark cited by Mistral |
Benchmark results depend on the model version, effort, sampling, tools, scaffolding and dataset release. Do not combine scores from different harnesses into a synthetic ranking. Use public evidence to choose candidates, then test those candidates on frozen tasks from the intended workflow.
Pricing and access
Mistral Large 4 is available as a public-preview API in Mistral Studio under the model IDs mistral-large-4 and mistral-large-4-0. Public preview means applications should pin a version when possible, monitor release notes and avoid assuming permanent production behavior.
Mistral announced that the model weights would follow after the hosted preview. Until the files, license and checksums are actually published, buyers should describe the model as open-weight planned rather than already downloadable. Self-hosting estimates should wait for the final artifacts and serving guidance.
Usage or access item | Current value |
|---|---|
Standard input | $1.36 per million tokens |
Cached input | $0.14 per million tokens |
Standard output | $4.18 per million tokens |
Launch promotion | Check Mistral billing for any time-limited preview discount |
Hosted access | Mistral Studio and Mistral API |
Weights | Announced as coming soon; not yet available when checked |
Token price is only part of deployment cost. Include prompt construction, cached-input behavior, tool calls, retries, wall time, failed-task recovery and human review. A cheaper model can be more expensive operationally when it needs repeated attempts or produces artifacts that require substantial correction.
Limitations and deployment risks
- The API remains a public preview and may change before general availability.
- The promised open weights were not yet published when this guide was checked.
- A one-million-token context window needs retrieval tests and explicit cost controls.
- Cybersecurity capabilities are dual use and require authorization, logging and target controls.
- Many launch benchmarks are provider-selected and use different harnesses or judges.
- The trillion-parameter footprint may require substantial infrastructure for self-hosting.
- Visual grounding and document claims should be tested on the exact image quality and document formats used in production.
High-stakes legal, medical, financial, scientific and security work needs qualified review. Log the exact model version and settings, separate untrusted content from system instructions, minimize tool permissions and require confirmation before irreversible or externally visible actions.
How to evaluate Mistral Large 4
Build a frozen evaluation set of 20 to 50 real tasks. Include routine work, difficult edge cases, adversarial inputs, long-context examples and cases where the correct behavior is to stop or escalate. Compare Mistral Large 4 with Mistral Large 3 and at least one neighboring model under equivalent conditions.
Test area | What to record |
|---|---|
Task completion | Pass or fail against a written acceptance rubric |
Reliability | Repeated-run success rate, variance and silent failures |
Factuality | Unsupported claims, source use, quotations and citation accuracy |
Tool use | Wrong calls, retries, recovery and permission-boundary failures |
Long context | Retrieval accuracy, instruction retention and cost at realistic lengths |
Efficiency | Wall time, input, output, cached tokens, tool charges and review time |
Safety | Prompt injection, sensitive-data handling and irreversible-action controls |
Choose the least expensive configuration that clears the acceptance threshold with a safety margin. Re-run the suite when the model, system prompt, effort, retrieval layer, tool definitions or approval policy changes.
Who should use it?
Situation | Recommendation |
|---|---|
Strong fit | multimodal enterprise agents, coding, cybersecurity, scientific work and European or self-hosted deployment planning |
Pilot first | Long-running agents, large contexts, computer use and workflows with several tools |
Escalate | Ambiguous or consequential work that does not reliably clear the evaluation threshold |
Avoid unsupervised use | Irreversible actions, sensitive data or high-stakes decisions without monitoring and approval |
A migration should be driven by measured outcomes. Keep Mistral Large 3 available during the pilot, record where each model succeeds or fails, and use routing when different task classes have different quality and cost requirements.
Frequently asked questions
Is Mistral Large 4 open source?
Mistral announced open weights, but they were not yet available when this guide was checked. The hosted public preview is available now. Verify the published license and files before describing a self-hosted deployment as production-ready.
How much does Mistral Large 4 cost?
$1.36 per million tokens. Review the full table because caching, prompt length, processing mode, platform and output volume can materially change the total.
What is Mistral Large 4 best used for?
Its strongest documented fit is multimodal enterprise agents, coding, cybersecurity, scientific work and European or self-hosted deployment planning. Start with a supervised pilot and retain human sign-off for consequential work.
Should I migrate from Mistral Large 3?
Only after a side-by-side evaluation. Measure accepted-task quality, total cost, latency, output style, tool reliability and migration engineering. Newer does not always mean better for a specific operational workflow.
How should benchmark evidence for Mistral Large 4 be interpreted?
Use provider results to identify promising workloads, then reproduce the comparison with the exact configuration, tools and acceptance criteria that matter to the deployment. Do not infer a site ranking from benchmark claims gathered under a different harness.
Related model guides
Browse the Best AI Models directory.
Compare with the Mistral Large 3 guide.
Compare with the DeepSeek V4 guide.
Compare with the Kimi K3 guide.
Official sources and update policy
Mistral Large 4 model documentation.
Checked October 22, 2026. We update this guide when Mistral AI changes the specification, pricing, access, safety documentation or lifecycle. Vendor benchmark claims are attributed and are not presented as independent testing. No vendor payment or affiliate relationship determined inclusion.