Mistral Large 4: Pricing, Open Weights, and When to Use It

Release, model cards, license, pricing and deployment documentation verified on 30 October 2026. Specifications below come from Mistral's released model cards; benchmark claims are attributed to Mistral, and the document budget is an illustrative calculation, not measured spending or a claim that this model has been tested here.

Mistral Large 4: What Changed With Open Weights

Since October 29, 2026, Mistral Large 4 can be downloaded and deployed on your own infrastructure. Mistral's open-weights announcement links the released model; its Hugging Face repositories provide the main FP8 checkpoint and the NVFP4 build. Buyers who require their own deployment can now validate an artifact and license instead of waiting for a promise.

The hosted public preview was announced on October 6, 2026.

The released cards describe approximately 1T total parameters and 49B active parameters per token, or 52B including embeddings and output layers, with a 1M-token context window. Mistral's API model card states the total more precisely as 1.05T. Parameters are the numerical values the model learns; a Mixture-of-Experts, or MoE, activates selected parts of the model for each token rather than using every parameter for every step. Active parameters describe computation, not the amount of model data your serving system must accommodate. Source: released model card.

The same weights handle instruction following and reasoning, with low, medium and high reasoning effort; high is the default. They accept text and images and produce text, with native function calling and strict JSON schema output. Structured output lets an application request a defined response shape; function calling lets it request an action from a tool your application exposes. These capabilities make charts, engineering drawings and evidence inside documents useful evaluation cases. Source: capabilities and settings.

The hosted route remains Mistral Studio and the API. The released card names mistral-large-4 as the API model ID. The API documentation lists chat completions, structured outputs, function calling and document question-answering. Use the model ID, endpoint and reasoning setting consistently when comparing hosted and local results.

Mistral Large 4 Price: API vs Self-Hosting

The API is the starting point when you need the model without operating its serving infrastructure. Mistral's own API pricing page now lists the following Standard rates in US dollars per million tokens, with regional inference switched off:

Tokens are the units the model processes: input includes your instructions and supplied material, while output is the generated material charged by the service. The two sides have different prices, so a single blended number can conceal an expensive answer-heavy workload. Our Mistral pricing guide carries the full model list and separate subscription options.

A Document Month Costs $24.50 Through the API

A stated workload of 1,000 document-review jobs costs $24.50 in Large 4 model charges at the current Standard rates. Assume each job averages 20,000 fresh input tokens and 5,000 billed output tokens, including every model call counted toward that job. The month therefore processes 20 million input tokens and 5 million output tokens. This is an illustrative usage budget, not a conversion from page counts or a measured deployment.

The calculation is:

• Input: 20 × $0.70 = $14.00.
• Output: 5 × $2.10 = $10.50.
• Total: $14.00 + $10.50 = $24.50.

These estimates exclude tax, credits, document-processing and tool charges, and any extra model usage beyond the stated token totals. The inputs come from Mistral's price sheet; the arithmetic is ours.

Output is only 20% of this example's token volume but about 43% of its model bill. Count the usage returned by the service across the whole task, rather than budgeting only the final paragraph the user sees.

The API Bill Sets the Hosting Comparison

Compare that usage bill with a complete operating plan, not the download price. Self-hosting adds capacity, upgrades, monitoring and failure recovery to your responsibilities. A lightly used deployment can carry idle capacity costs; a busy one can spread them across more tasks. Neither situation establishes that self-hosting will be cheaper than the API.

For a CTO, the decision can still favor self-hosting when control over the execution environment is mandatory. Budget the infrastructure and the people needed to operate it, then compare cost per successfully completed task at your required context length and concurrency. The cards do not supply a hosting quote or a universal break-even point.

The Cheaper Model Sets the Adoption Hurdle

A larger model earns its route through fewer rejected answers, less correction work or a required deployment boundary. Leave successful inexpensive routes alone. A low token price alone does not establish low cost per accepted result, and access to downloadable weights does not make every workflow worth migrating.

What the Benchmarks Establish

Mistral's results justify an evaluation shortlist; they do not establish a winner for your workload. A benchmark is a defined set of tasks run under particular conditions. Coding-agent scores involve tools and execution environments, not just a model answering a short prompt. Those distinctions decide whether a result is relevant to your application.

The following are Mistral's reported results for the released model, taken from its open-weights announcement. They are not results produced by this site, and percentages from different tests are not interchangeable.

Source for all table entries: Mistral's release evaluation. A security benchmark score is not a guarantee that an agent will resist every malicious document it encounters. Use the relevant workload to select a pilot, then check each result against the supplied evidence.

Read the Scores as Workload Signals

The released checkpoint needs its own evaluation under your serving configuration. A benchmark claim does not establish equal behavior at every reasoning setting, context length or concurrency level. Hold the evidence, tools and acceptance criteria consistent when you compare the API with a local deployment. There is no site-run ranking here: the decision rests on verified deployment terms and checks you can apply to your own work.

Mistral Large 4 Review: Who Should Test It This Month

Test Large 4 when a named failure in your current route matters enough to pay for a better accepted result. Leave successful inexpensive routes alone. That gives builders, operators and enterprise buyers different reasons to act, even though they are looking at the same release.

Builders: Add a Route for the Difficult Cases

A funded founder should pilot one difficult workflow before changing the default model. For example, a document assistant may handle ordinary field extraction cheaply but struggle when evidence spans charts, footnotes and conflicting exhibits. Put that subset into the evaluation and measure whether Large 4 reduces corrections. Its multimodal and tool-use claims make the question reasonable; only your acceptance results decide the route.

A solo technical builder whose existing model already returns acceptable classifications has little reason to switch. Keep the smaller route and retain Large 4 as a candidate for a task that fails it. For agent-model alternatives, the compact coding models guide gives a narrower starting point than replacing every call with a new flagship.

Operators: Security and Evidence-Heavy Documents Come First

A security lead has a concrete reason to test authorized defensive work. Mistral names malware analysis, vulnerability prioritization and detection-rule writing, alongside vulnerability reproduction and patching. A useful pilot asks whether the proposed patch addresses the supplied issue and whether the resulting detection rule matches the evidence. Keep the allowed actions explicit and review the output before it changes operational systems. Source: Mistral's cybersecurity section.

For a legal operations lead, use an approved document set and score whether an answer identifies the controlling passage, preserves exceptions and distinguishes evidence from inference. For a finance operator, score whether the answer uses the right period, reproduces the calculation and traces each input to the source document. A polished memorandum with the wrong clause or period is rejected work, regardless of the benchmark behind the model.

Mistral also names engineering drawings, manufacturing, geospatial imagery and scientific programming. These are separate evaluation tracks: identifying a component in a technical drawing, finding an object in satellite imagery and implementing a scientific routine need different acceptance criteria. The release makes them candidates; it does not prove one general score transfers to all three. Source: Mistral's capability deep-dive.

Enterprise Buyers: Verify the EU Boundary You Are Buying

EU data-location requirements are a reason to compare a regional API with your own deployment, then verify the exact service boundary. Mistral's general regional-inference documentation says the global endpoint does not commit to a specific processing location. Choosing a European model developer alone does not establish where every request is processed. Source: regional inference.

The regional route is api.eu.mistral.ai. Check that the required model is actually available there, because regional model availability varies. Its region table describes datacenters in EU and EFTA countries; EFTA is the separate European Free Trade Association. A strictly EU-only requirement therefore needs confirmation of the permitted processing locations. Mistral also says regional endpoints do not provide Agents, Batch or the Files API, and function calling is the only supported regional tool. Source: regional features.

That matters for a CTO planning a document agent: the global model card's feature list does not establish that the same managed workflow runs through the regional endpoint. Request-and-response processing geography also does not regionalize every account, billing or usage record, and it is separate from zero data retention. Verify the processing, storage and tool dependencies your particular system needs before moving regulated material.

What the Released Weights Change

The weights give you a deployment path now, with a license and a substantial serving job attached. Open weights are downloadable model files. The main checkpoint and NVFP4 checkpoint let a buyer validate the execution environment, access boundary and operating procedures before approving production use.

The License Allows Commercial Use With a Hosting Restriction

Both builds use the Mistral Ecosystem Open-Weight License, a modified MIT license. It permits commercial and non-commercial use subject to its conditions. It is not an Apache 2.0 release or an unrestricted MIT grant. Sources: main license, NVFP4 license.

If you or an affiliate operates a Model-as-a-Service business, the license requires a separate commercial agreement with Mistral before commercial use, with exceptions for internal use and use through Mistral's official products. Its service definition covers third-party access with meaningful control over inputs, parameters or training data; it excludes capabilities embedded solely within specific end-user product features and merely relaying requests to models hosted by others. An internal document assistant and a general-purpose model endpoint for customers therefore raise different licensing questions. Our open-weight models guide explains the distinction between downloadable weights and deployment rights.

What Self-Hosting Takes

Plan for multi-GPU model serving, not a model that fits like a 52B download. Mistral's main card recommends vLLM and documents SGLang and Transformers. Its SGLang recipe explicitly describes an 8-GPU node, and its Transformers example also runs on 8 GPUs. These are documented examples, not proof that any server with that GPU count can serve the full context window at your required concurrency.

The card tells operators who run short of GPU memory to reduce --max-model-len, the served context length, and/or --max-num-seqs, the number of concurrent sequences. That is the practical constraint: fitting the weights does not by itself establish capacity for long requests and simultaneous users. Neither card gives a universal minimum GPU-memory capacity or guaranteed throughput for your workload.

For an enterprise buyer, ownership also moves responsibility for upgrades, monitoring and failure recovery to whoever operates the deployment. If you cannot already support that serving environment, keep the API as the default while you validate the infrastructure. If control over where inference runs is mandatory, the release now lets you start that validation.

What the NVFP4 Build Changes

NVFP4 changes the numeric format and hardware target, not the model's deployment responsibilities. Quantization stores model values at lower precision. This build uses NVFP4 MoE experts with FP8 linear layers and targets NVIDIA Blackwell B200, B300, GB200 and GB300 GPUs. Mistral claims a substantially smaller memory footprint and faster inference with similar performance to FP8; that is a vendor claim to validate on your workload, not a promised speedup. Source: NVFP4 model card.

The NVFP4 card retains the 1M context window, text-and-image input, reasoning settings, function calling and strict JSON output. Its vLLM example sets --tensor-parallel-size 8; its SGLang example uses --tp-size 4, described as an example for a Blackwell node to adjust to your topology. Those different recipe settings are not a universal minimum hardware specification. The vLLM recipe also requires its documented fp8_ds_mla KV-cache setting, which controls the format used for cached attention state. Follow the checkpoint-specific recipe rather than substituting a quantized filename into an unrelated configuration.

Use the best open-source LLMs comparison to compare alternatives if this infrastructure commitment is larger than your workflow warrants. Choose self-hosting because the control and operating plan make sense, not because the weights are downloadable.

What Is Overhyped

Downloadable weights and a production deployment are different milestones. The artifact now exists. A deployable, licensed, supportable system still needs its own validation: the license must fit your product, the serving recipe must fit your hardware, and the workload must meet its acceptance criteria.

The benchmark charts also do not establish universal superiority. Security refusals, tool access, execution environments and task definitions can affect scores. A spreadsheet benchmark does not settle whether your reporting workflow preserves the right evidence. Use the relevant claim to select a test, then score the deliverable you need.

Expanded cybersecurity access is another boundary. Mistral's release post describes a specialized cyber variant for vetted enterprise design partners, with contractual safeguards and testing restricted to owned or explicitly authorized systems. Do not assume the public weights or a standard API account provide the same variant. Source: specialized cyber access.

Finally, reduced weight precision does not establish lower total operating cost. Compare the actual deployment boundary, accepted results and operating effort before migrating.

Your Next Move: A Bounded Pilot

Next week, an engineering lead should run the same approved evaluation set through the API and, if deployment control is required, the released checkpoint on a card-supported serving setup. Keep production traffic on the current accepted route until the candidate passes. An enterprise buyer can now review the actual license and serving plan.

For a basic global-endpoint smoke check, with your own Mistral API key already set in MISTRAL_API_KEY, the documented chat-completions route is:

curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-large-4",
    "messages": [
      {"role": "user", "content": "Explain the difference between input and output tokens in two sentences."}
    ]
  }'

That checks connectivity and model selection, not document quality or regional compliance. The adoption decision comes from the approved workload and its accepted results.

Get practical model and workflow decisions in the newsletter.