Mistral Large 4 Opens Public Preview With a 1.05T MoE
Mistral’s flagship LLM combines million-token context, image understanding, and discounted API access, while its open-weight release remains a separate milestone.
Listen
AI narration
12:05
0:00 / 12:05
AI SummaryGenerated from this article
Mistral Large 4 is available through public preview API with a 1.05 trillion parameter Mixture-of-Experts model, but its downloadable weights remain unreleased. The model features 52 billion active parameters, a 1.6 billion parameter vision encoder, one million token context window, and multimodal capabilities for documents and images.
The preview API is offered at 50% discount for two weeks, with input tokens at $0.68 per million and output at $2.09 per million. Developers should verify performance on their specific documents and workloads before production deployment, as the preview status and long-context accuracy remain unproven at scale.
Mistral Large 4 is available through a public preview API, but its downloadable weights are not part of the initial release. That distinction matters for developers evaluating both the model’s capabilities and Mistral’s promise of deployment independence.
The Large 4 model card describes a multimodal Mixture-of-Experts model with 1.05 trillion total parameters, 52 billion active parameters, a 1.6-billion-parameter vision encoder, and a one-million-token context window. Those specifications make this a substantial flagship model release, although they do not establish how reliably it will perform in a particular application.
For developers, the immediate opportunity is concrete: test the hosted model’s document understanding, structured outputs, and tool use at discounted launch prices. The decision to put it into production requires a separate assessment of preview stability, long-context accuracy, operating costs, and deployment terms.
The API Is Available Before the Weights
Mistral’s October 6 changelog identifies mistral-large-4 as available in public preview. The documentation labels the release “Public Preview” and “Open v26.10,” rather than generally available.
The company’s Large 4 announcement says developers can try the preview API in Mistral Studio and that weights will arrive by the end of October. The changelog likewise says open weights are “coming soon.”
That establishes two different milestones:
Available at launch: hosted preview access through Mistral’s API.
Announced for later: downloadable weights for deployment outside that hosted service.
Calling Large 4 an open-weight model therefore describes Mistral’s release commitment, not evidence that developers can already download and run it. Teams planning private-cloud or on-premises deployments should verify the actual weight publication, license, supported inference software, and hardware requirements before making those plans operational.
The announcement also says Mistral is still refining the model and conducting real-world red-teaming before releasing the weights. Certain vetted partners receive reduced moderation and expanded cybersecurity capabilities. Developers using the ordinary public preview should not assume their access has the same behavior or permissions.
A Trillion Parameters Does Not Mean a Trillion Active
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
Large 4’s headline number describes its total parameter count. Its Mixture-of-Experts, or MoE, architecture uses only a subset of its expert parameters during computation, with Mistral reporting 52 billion active parameters.
That active count is approximately 5% of the stated 1.05-trillion total. Sparse activation gives the model access to a large pool of learned parameters without activating the entire pool for every token.
The distinction helps explain how a trillion-parameter model can be served economically. It does not, however, justify treating Large 4 as equivalent to a dense 52-billion-parameter model. Routing, memory movement, expert placement, and the rest of the inference system affect the cost and speed of serving an MoE.
Nor does sparse activation eliminate the storage and memory burden of the full weights. A self-hosting plan must account for the model’s total size and serving architecture, not just the active-parameter figure.
Mistral describes the design as a “granular” MoE, but the announcement says further architectural and post-training details will follow. The available specifications do not establish the exact routing configuration or a validated minimum hardware setup.
The 1.6-billion-parameter vision encoder is another relevant specification: Large 4 is built to process visual information, rather than being presented solely as a text LLM. Its size alone does not demonstrate visual accuracy.
Multimodal Understanding and Tool Use Have Different Boundaries
Mistral describes Large 4 as natively multimodal and highlights understanding of images, complex documents, charts, and technical drawings. The announcement includes visual-grounding and document-analysis demonstrations.
The supported conclusion is that Large 4 offers text-and-image understanding. The cited material does not establish native audio processing, video support, or image generation, so “multimodal” should not be interpreted as support for every input and output format.
Structured outputs, function calling, document Q&A, and prefix support
/v1/chat/completions, /v1/conversations
Chat completions
/v1/chat/completions
Batching
/v1/batch
Agents and conversations
/v1/agents, /v1/conversations
Built-in tools
/v1/agents, /v1/conversations
These are integration details, not interchangeable labels. Function calling allows an application to expose functions that the model can request. Built-in tools belong to Mistral’s agent and conversation interfaces, according to the feature matrix.
A developer migrating an existing chat-completions application should therefore check which capabilities work directly and which require a different API workflow. Support for tool use does not mean every tool is available through every endpoint.
The same distinction applies to document Q&A. Listing that feature does not guarantee accurate extraction from every scanned page, chart, or engineering drawing. Teams should test their own document types, including cases where the correct answer depends on small visual details or requires admitting that evidence is missing.
One Million Tokens Is a Capacity Claim
Large 4’s published one-million-token context window makes large-document and repository-scale experiments possible within a single request. It is a capacity specification, not a guarantee that the model will retrieve every relevant detail or reconcile every contradiction correctly.
Long-context evaluation should separate several questions: whether a request fits, whether the model finds the needed evidence, and whether its answer follows from that evidence. A successful short-document test does not settle performance near the context limit.
Developers should check how text, images, tool definitions, conversation history, and reserved output capacity count toward the available window. The supplied model card does not resolve those accounting details.
Useful tests would place decisive evidence at different positions, introduce conflicting statements, and include questions that the supplied material cannot answer. These expose failures that a simple “find this fact” demonstration can miss.
Latency matters too. A context window large enough to accept a workload may still produce response times unsuitable for an interactive application. Compare Large 4 against alternatives using the same documents, task, and response-time requirements, rather than comparing context limits alone.
The Displayed Prices Are a Two-Week Launch Discount
Mistral’s documentation currently displays the following USD prices per million tokens:
Token category
Displayed sale price
Listed original price
Input
$0.68
$1.36
Cached input
$0.07
$0.14
Output
$2.09
$4.18
The changelog describes the offer as 50% off for two weeks. These are launch prices, not an established long-term price commitment. Developers budgeting beyond the promotion should use the listed original rates as a planning reference and confirm the applicable prices before deployment.
For an illustrative request with 100,000 uncached input tokens and 5,000 output tokens, the listed token charges total approximately $0.078 at sale rates, compared with $0.157 at the original rates. That calculation covers one request’s input and output, not repeated agent calls or retries.
The cached-input rate could materially lower costs for repeated eligible content. A listed cache price does not mean that every repeated prompt receives it, however. Teams should verify cache eligibility and inspect actual usage accounting.
For agentic applications, cost per completed task is more useful than cost per initial request. An inexpensive model call can become expensive if the workflow repeatedly searches, invokes tools, repairs invalid arguments, or restarts after an error.
Production Readiness Requires Evidence Beyond the Preview
Mistral makes strong performance claims across coding, cybersecurity, visual grounding, finance, and law. Those claims explain the company’s positioning, but the supplied evidence does not independently reproduce the broader results. They should guide evaluation priorities rather than substitute for application-specific testing.
Before routing important workloads to Large 4, developers should verify four areas.
Model identity and lifecycle. Confirm whether the selected identifier is version-pinned, what changes can occur during preview, and what notice or migration arrangements apply. “Public preview” is not, by itself, evidence of a stable production lifecycle.
Operational behavior. Measure throughput, time to first token, full-response latency, rate limits, and error handling under realistic concurrency. Include long prompts and tool-heavy sessions. Check what availability commitments apply to the actual account and service tier.
Tool and output reliability. Test schema adherence, argument validity, ambiguous instructions, and recovery after failed tool calls. Keep authorization and validation in application code. A model’s request to perform an action should not automatically authorize that action.
Data and deployment terms. Mistral says the preview runs on its own European infrastructure, but that statement alone does not resolve every customer’s retention, processing, or residency requirements. Verify the applicable contractual terms, region, logging behavior, and data-handling settings.
Large 4’s release is meaningful because developers can now evaluate a sparse trillion-parameter model through a public API without first building the infrastructure to host it. Its strongest practical test is whether it completes real workloads accurately, within the required latency, at an acceptable total cost.
For teams whose main requirement is control over deployment, the decisive milestone is still ahead: published weights with usable licensing and a supported serving path. Until then, the preview provides access to evaluate the model, not a completed route to independent production deployment.
