Mistral AI Models: Every Model, Spec, Price and Release Date

Mistral AI is a French AI company that publishes large language models, including open-weight models, through its La Plateforme API and on Hugging Face; its current flagship, Mistral Large 4, was released as a public preview on 6 October 2026 and reached general availability on 5 November 2026. This page lists every current Mistral model, its API identifier, context window, licence, published price and status, plus the version history, the deprecation and retirement schedule and an index of model cards. Every fact comes from Mistral's own documentation, changelog, model cards or pricing page. Where Mistral publishes no value, the page says "Not stated by Mistral AI" rather than estimating. Facts are as of 15 November 2026.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Mistral Large 4 is the flagship. Version 26.10, API names mistral-large-4, mistral-large-4-0 and mistral-large-latest, a 1M token context length and a Mixture-of-Experts design with 1.05T total and 52B active parameters, according to its model card.
• Most generalist models carry open licences. Mistral Large 4 uses the Mistral Ecosystem Open Weight License; Mistral Small 4 and the Ministral 3 family use Apache 2.0.
• Two official prices exist for "Large". The Large 4 model card lists $0.70 input, $0.07 cached input and $2.10 output per million tokens; the general pricing page shows "Mistral Large" at $0.5 and $1.5. This page uses the more specific model card figure and notes the other.
• A large retirement wave is scheduled. Mistral Large 3, Mistral Medium 3.5 and several Voxtral models were deprecated on 31 October 2026 and retire on 30 April 2027.
• The catalogue goes beyond chat models. Mistral also lists OCR, speech (Voxtral), code (Codestral), embedding, moderation (Mistral Moderation 2, Shieldstral) and time series (Timestral) models, and hosts third-party Z.ai GLM and DeepSeek models.

What is Mistral AI?

Mistral AI publishes models under two broad licence groups in its models overview: open models (Apache 2.0, Modified MIT or the Mistral Ecosystem Open Weight License, with weights published) and "Premier" models available through the API. Each model carries a version label in the form year.month, so v26.10 means October 2026 and v25.12 means December 2025. API identifiers often encode the same date, for example mistral-small-2603 for Mistral Small 4 (March 2026). Floating aliases such as mistral-large-latest point to the current release. See the LLM Model Naming and Versioning Explained page for how these conventions work across providers.

The models overview lists 24 current models and a long list of deprecated or retired ones as of 15 November 2026. Models are grouped below as Mistral groups them: generalist, OCR, audio, code, embedding, moderation and safety, and time series.

Current Mistral models at a glance

As of 15 November 2026. Source: Mistral models overview and individual model cards. Models are listed in Mistral's own category order.

Status for the Voxtral 3 models comes from the changelog entry of 20 October 2026, which lists them as public preview. GLM and DeepSeek models are developed by Z.ai and DeepSeek respectively and are served on Mistral's platform; their own provider pages are the primary source for model details.

Key facts for the main Mistral models

As of 15 November 2026. Identical fields for every model; "Not stated by Mistral AI" means not stated by Mistral AI on the pages checked (2026-11-15). API IDs, context, licence and status are in the table above.

Status, API IDs and context for Mistral Medium 3.5, which is deprecated and so not in the table above: status deprecated 2026-10-31, API ID mistral-medium-3-5, context 256k, licence Modified MIT. Ministral 3 status is GA on the 14B card and "current" on the overview; API IDs are ministral-14b-2512, ministral-8b-2512 and ministral-3b-2512. Mistral Large 4 API names are mistral-large-4, mistral-large-4-0 and mistral-large-latest; Mistral Small 4 is mistral-small-2603.

Mistral Large 4 and Mistral Small 4 both use a Mixture-of-Experts design, where only part of the network is active per token, which is why each card gives a total and an active parameter count. See Mixture of Experts Explained.

API features listed on the model cards

Mistral model families and naming

Mistral AI groups its models into the lines below, as of 15 November 2026. Release dates for each are in the next section.

• Large line: Mistral Large 1.0 (24.02), Large 2.0 (24.07), Large 2.1 (24.11), Large 3 (25.12) and Large 4 (26.10). The Medium 3.5 model card names Mistral Large 4 as its replacement for new integrations.
• Medium line: Medium 1.0 (23.12), Medium 3 (25.05), Medium 3.1 (25.08) and Medium 3.5 (26.04), all now deprecated.
• Small line: Small 1.0 (24.02), 2.0 (24.09), 3.0 (25.01), 3.1 (25.03), 3.2 (25.06) and Small 4 (26.03).
• Reasoning and coding lines folded in: Magistral (reasoning) and Devstral (coding) models are deprecated; Mistral Small 4's card describes it as unifying instruct, reasoning and coding in one model. See Reasoning Models Explained.
• Edge line: Ministral 3B and 8B (24.10, retired) were followed by Ministral 3 in three sizes (25.12). The 14B card cites a technical report, arXiv 2601.08584.

When was each Mistral model released?

Dates from the Mistral changelog and models overview, as of 15 November 2026, in date order.

Mistral Large 4 entered public preview on 6 October 2026 and its model card gives a release date of 5 November 2026 with GA status. The most recent changelog entries are the Voxtral 3 public previews (20 October 2026) and Z.ai GLM 5.3 Flash on La Plateforme (22 October 2026). The changelog had no November 2026 entries when checked.

How much do Mistral models cost?

As of 15 November 2026. USD per million tokens. Sources: individual model cards and the Mistral pricing page (mistral.ai/pricing).

Prompt caching: Mistral's prompt caching documentation states that cached prompt tokens are billed at 10% of the standard input token price, that cache blocks contain 64 tokens, and that prompts under 64 tokens get no cache hits. A prompt_cache_key parameter raises the chance of a hit but does not guarantee one. No separate cache write price is stated.

Conflict note: the Mistral Large 4 model card and the general pricing page give different figures for "Large". The model card is the more specific official page for mistral-large-4, so this page uses $0.70 input, $0.07 cached and $2.10 output; the pricing page's $0.5 and $1.5 for "Mistral Large" is shown for completeness. Confirm the price in the console before committing budgets. The model cards show both USD and EUR toggles.

Consumer and team plans

For how input, output, cached and batch pricing work in general, see LLM Pricing Explained: Tokens and Caching. To cap spend on keys, see How to set spend limits on LLM API keys.

Where are Mistral models available?

• First-party API: La Plateforme, with endpoints including /v1/chat/completions, /v1/conversations, /v1/agents and /v1/batch, as listed on the model cards.
• Open weights: Mistral Large 4 weights are linked from its model card at huggingface.co/mistralai/Mistral-Large-4-1T-A52B. The Small 4 card shows a Weights tab, and the Medium 3.5 card and changelog state open-weight release; the direct links were not shown in the text checked.
• Apps: the pricing page lists consumer and team plans (Free, Pro, Team, Enterprise) under the product name Vibe.
• Third-party clouds and regions: not stated by Mistral AI on the pages checked for this update.

Running open weights yourself moves security responsibility to you. See the Open-Weight Model Security Guide and the Self-Hosted LLM Security Guide.

Restricted-access Mistral models

As of 2026-11-15. Restricted-access programmes for Mistral models: not stated by Mistral AI on the pages checked (2026-11-15).

Mistral's zero data retention page states that ZDR applies to supported endpoints across models except Labs models. Hosted third-party models (Z.ai GLM and DeepSeek) are listed as current on La Plateforme; Mistral states no access restriction for them on the pages checked.

What is the context window of Mistral models?

Mistral Large 4 has a 1M token context length on its model card. Mistral Small 4, Ministral 3 14B and Mistral Medium 3.5 have 256k. Mistral Moderation 2 has 128k. The hosted GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash models are listed at 1M. See LLM Context Windows Explained.

Is any Mistral model open source?

Several Mistral models are open weight. Mistral Small 4, Ministral 3 and Voxtral Small use Apache 2.0. Mistral Medium 3.5 and Timestral use a Modified MIT licence. Mistral Large 4 uses the Mistral Ecosystem Open Weight License, a Mistral-specific licence, so check its terms before commercial use. OCR, Codestral, embedding, Voxtral 3 and Mistral Moderation 2 are listed as Premier (API) models. See Open-Weight vs Closed LLMs Explained.

Model cards and system cards

Mistral's model cards are short specification pages (release date, version, API names, context, parameters, features, price, licence). They do not include safety evaluation sections on the pages checked. See Model Cards and System Cards Explained.

Benchmarks

As of 15 November 2026. Provider-reported and independent results are kept in separate columns.

See How to Read LLM Benchmarks for how provider-reported and leaderboard figures differ.

Deprecations and retirements

As of 15 November 2026, from the models overview and changelog. Upcoming and recent dates only; the overview lists older retirements back to 2024.

Pinned dated identifiers stop working at retirement, while aliases such as mistral-large-latest move to the new model. Inventory where each identifier is used before the date. See the AI Model Retirement Guide.

Data handling and security

• Training use: Mistral's help centre states that API and Studio users can opt out of their data being used for training in the Admin panel, under Privacy, by disabling "Anonymous improvement data". Whether API data is used for training by default is not stated by Mistral AI on the pages checked (2026-11-15). For the Vibe app, Enterprise customers are opted out by default, Team administrators can disable training for the organisation, and individual users are not opted out by default. The Vibe and API toggles are separate.
• Default retention: not stated by Mistral AI on the pages checked (2026-11-15).
• Zero data retention (ZDR): requested through Mistral support with a stated reason; Mistral reviews each request and may deny it. If approved, ZDR appears under Admin Panel, Vibe, Privacy. With ZDR, inputs and outputs for supported API requests are not stored or logged longer than needed to generate the output.
• Not ZDR-eligible: Labs models and stateful APIs and products, including Agents, Batch processing files, Conversations, Libraries, /v1/files, Vibe Work, Chat and traces sent to Studio Observability.
• Key handling: generic controls apply to every provider's API: keep keys out of code, scope and rotate them, and set spend limits.

Sources: Mistral zero data retention page (docs.mistral.ai/admin/monitor-comply/zero-data-retention) and help centre article on training opt-out, both accessed 2026-11-15. See the LLMjacking Guide and the Foundation Model Selection Guide.

How to choose a Mistral model: factual criteria

• Licence: Apache 2.0 (Small 4, Ministral 3, Voxtral Small), Modified MIT (Medium 3.5, Timestral), the Mistral Ecosystem Open Weight License (Large 4) or Premier (API only).
• Self-hosting: open-weight models can be run on your own infrastructure; Premier models are API only.
• Price per token: compare the model card and pricing page figures, which differ for "Large"; cached tokens are 10% of input.
• Context: 1M on Large 4, 256k on Small 4 and Ministral 3 14B, 128k on Moderation 2.
• Retention requirements: ZDR excludes Labs models and stateful APIs such as Agents, Conversations and Batch files.
• Lifecycle: Large 3 and Medium 3.5 retire on 30 April 2027; check the stated replacement before building.
• Features: check the model card feature list, since predicted outputs, built-in tools and connectors differ by model.

Practitioner checklist

• Confirm the exact model ID and version on Mistral's current models page before you build against it.
• Check the price per million input and output tokens on the official pricing page and the model card, and note the date you checked.
• Check the context window and maximum output for the specific model, not the model family.
• Read the deprecation and retirement notices and plan migration dates for any model you depend on.
• Review Mistral's data handling, retention and training terms for the plan or platform you will use.
• Store provider API keys in a secrets manager and rotate them; never embed them in client code.
• Plan the move off mistral-large-2512 and mistral-medium-3-5 before 30 April 2027.
• Check whether any workload uses a Labs model or a stateful API before relying on ZDR.

Standards and references

• Mistral AI: Models overview (accessed 2026-11-15)
• Mistral AI: Mistral Large 4 model card (accessed 2026-11-15)
• Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
• Mistral AI: Ministral 3 14B model card (accessed 2026-11-15)
• Mistral AI: Mistral Medium 3.5 model card (accessed 2026-11-15)
• Mistral AI: Changelog (accessed 2026-11-15)
• Mistral AI: Pricing (accessed 2026-11-15)
• Mistral AI: Zero data retention, docs.mistral.ai/admin/monitor-comply/zero-data-retention (accessed 2026-11-15)
• Mistral AI: Prompt caching, docs.mistral.ai/studio-api/conversations/advanced/prompt-caching (accessed 2026-11-15)
• Mistral AI Help Center: Can I opt out of my input or output data being used for training? (accessed 2026-11-15)

Related NHI Mgmt Group resources: xAI Models, LLM Security Guide, Open-Weight Model Security Guide, Foundation Model Selection Guide.