Join our Newsletter — 33% off our NHI Course
Home› Guides› Mistral Large 4 vs Medium 3.5 vs Small…
Guide AI Platforms

Mistral Large 4 vs Medium 3.5 vs Small 4: Specs, Prices and Lifecycle Side by Side

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 15 November 2026 17 min read
On this page

Mistral AI has three recent general-purpose models in its documentation, Mistral Large 4, Mistral Medium 3.5 and Mistral Small 4, and they differ on context window, architecture, licence, price and lifecycle status. This page compares the three side by side: API IDs, release dates, status, context window, parameter counts, prices, reasoning controls, API features, availability, provider-reported benchmarks and retirement dates. It is written for developers choosing or migrating between Mistral models, buyers estimating cost and architects who need to know which ID will still be callable next year. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Mistral AI's own documentation, model cards and announcements, cited in the sources list.

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. It uses Mistral AI's public documentation as of November 2026. Model details change, so confirm against Mistral AI's current documentation.

Key takeaways

  • Two models are current and one is deprecated. Mistral's models overview lists Large 4 (26.10) and Small 4 (26.03) as current; Medium 3.5 (26.04) was deprecated on 31 October 2026 and retires on 30 April 2027.
  • Medium 3.5 points to Large 4. The Medium 3.5 model card states "Use Mistral Large 4 for new integrations" and names mistral-large-4-0 as the replacement.
  • Context runs from 256k to 1M tokens. Medium 3.5 and Small 4 list 256k tokens; Large 4 lists 1M tokens. None of the three model cards states a separate maximum output.
  • Listed prices differ by more than tenfold on output. Per million tokens: Large 4 $0.70 input and $2.10 output, Medium 3.5 $1.50 and $7.50, Small 4 $0.15 and $0.60, each from its own model card.
  • All three are open weight, under three licences. Large 4 uses the Mistral Ecosystem Open Weight License, Medium 3.5 a Modified MIT licence and Small 4 Apache 2.0, so self-hosting terms are not interchangeable.

What are Mistral Large 4, Medium 3.5 and Small 4?

They are the latest releases in Mistral AI's Large, Medium and Small lines, each of which accepts text and images and returns text with an optional reasoning trace. Mistral numbers versions as year.month, so 26.10 means October 2026; see Mistral AI Models for the full line-up and naming.

  • Mistral Large 4: a Mixture-of-Experts model with 1.05T total and 52B active parameters plus a 1.6B vision encoder (model card). The launch post calls it a 1-trillion-parameter, natively multimodal model trained from scratch on European infrastructure, nicknamed "Le Chonk".
  • Mistral Medium 3.5: a dense 128B model (Hugging Face card) that Mistral describes as "Our frontier-class multimodal model optimized for agentic and coding use cases". The Hugging Face card calls it a merged model handling instruction following, reasoning and coding in one set of weights.
  • Mistral Small 4: a Mixture-of-Experts model with 119B total and 6.5B active parameters (model card). The launch post says it combines the capabilities of Magistral (reasoning), Pixtral (multimodal) and Devstral (agentic coding) in one model.

Mixture-of-Experts models activate only part of the network per token, so total and active counts both matter for hosting; see Mixture of Experts Explained. Deep detail on two of the three models is in the Mistral Large 4 and Mistral Small 4 profiles; this page does not repeat them.

Every model at a glance

As of 2026-11-15. Sources: each model's documentation card, Hugging Face card and Mistral's models overview.

ModelAPI IDs listed by MistralVersionReleasedStatusParametersLicence
Mistral Large 4mistral-large-4, mistral-large-4-0, mistral-large-latest26.10Public preview 2026-10-06; model card release date 2026-11-05GA, current1.05T total, 52B active, plus 1.6B vision encoder (MoE)Mistral Ecosystem Open Weight License
Mistral Medium 3.5mistral-medium-3-5, mistral-medium-3, mistral-medium-latest26.042026-04-28Deprecated128B (dense)Modified MIT
Mistral Small 4mistral-small-2603 (the card shows one further alias without naming it)26.032026-03-16GA, current119B total, 6.5B active (MoE)Apache 2.0

Open weights: yes for all three. Hugging Face repositories named by Mistral: mistralai/Mistral-Large-4-1T-A52B, mistralai/Mistral-Medium-3.5-128B and mistralai/Mistral-Small-4-119B-2603. Knowledge cutoff: not stated by Mistral AI for any of the three on the pages checked (2026-11-15). The Small 4 launch post gives 6B active parameters (8B including embedding and output layers), while the model card gives 6.5B; this page uses the card.

When was each model released?

As of 2026-11-15. Dates from the Mistral changelog, models overview and model cards.

DateEventAPI ID
2025-12-02Mistral Large 3 released alongside the Ministral 3 modelsmistral-large-2512
2026-03-16Mistral Small 4 released as a hybrid instruct, reasoning and coding model with a 256k windowmistral-small-2603
2026-04-28Mistral Medium 3.5 released as open weights under a Modified MIT licence, with adjustable reasoning through reasoning_effortmistral-medium-3-5
2026-05-22Mistral Medium 3 and 3.1, Magistral Medium 1.2 and Devstral 2 deprecated, replacement Medium 3.5mistral-medium-2505, mistral-medium-2508, magistral-medium-2509, devstral-2512
2026-08-31Mistral Medium 3 and 3.1 retiredmistral-medium-2505, mistral-medium-2508
2026-10-06Mistral Large 4 in public preview, with launch pricing 50% off for two weeks and open weights to followmistral-large-4
2026-10-31Mistral Medium 3.5 and Mistral Large 3 deprecated, replacement Large 4mistral-medium-3-5, mistral-large-2512
2026-11-05Mistral Large 4 model card release date, status GA, weights on Hugging Facemistral-large-4-0
2027-04-30Scheduled retirement of Mistral Medium 3.5 and Mistral Large 3mistral-medium-3-5, mistral-large-2512

Medium 3.5 had a short life as a current model: just over six months from release to deprecation. Its own replacement role was large, since Mistral's overview names it as the successor to Medium 3 and 3.1, Magistral Medium 1.0 to 1.2 and Devstral 2.

How much does each model cost?

As of 2026-11-15. USD per million tokens, from each model's documentation card unless noted.

ModelInputCached inputOutputBatchNotes stated by provider
Mistral Large 4$0.70$0.07$2.1050% reduction (pricing page, general)Launch post repeats the card figures; pricing page FAQ quotes "Mistral Large" at $0.5 input and $1.5 output
Mistral Medium 3.5$1.50Not stated by Mistral AI on the pages checked (2026-11-15)$7.5050% reduction (pricing page, general)Deprecated; prices from the Markdown version of the model card
Mistral Small 4$0.15Not stated by Mistral AI on the pages checked (2026-11-15)$0.6050% reduction (pricing page, general)The card lists the two figures without labels; the launch post labels them input and output

Cached input in general: Mistral's pricing page FAQ states that cached input tokens reduce input cost by up to 90% for repeated prompts, without per-model figures. Large 4's $0.07 is 10% of its input rate.

Price conflict for Large: two official Mistral pages give different figures for "Large". The model card is specific to mistral-large-4, so this page uses $0.70 and $2.10. Check the rate shown in Mistral Studio before committing a budget.

Self-hosting: with open weights there is no per-token fee from Mistral, only your own compute. Large 4's licence requires a separate commercial licence to run a Model-as-a-Service business, and the Medium 3.5 Hugging Face card describes its Modified MIT licence as having "exceptions for companies with large revenue". For cost controls, see How to set spend limits on LLM API keys and the LLM Pricing Tracker.

What changed between models?

Only changes Mistral documents in its model cards, changelog and announcements are listed. The three models are in different lines, so this records how each relates to its predecessors and to the others.

  • Small 4 (March 2026): brought reasoning, vision and agentic coding into the Small line. The launch post says it unifies the Mistral Small instruct line with Magistral (reasoning), Pixtral (multimodal) and Devstral (agentic coding). reasoning_effort takes none (replies in the style of Small 3.2) or high (step-by-step reasoning with verbosity comparable to earlier Magistral models).
  • Medium 3.5 (April 2026): the Hugging Face card says it replaces Mistral Medium 3.1 and Magistral in Le Chat and Devstral 2 in the Vibe coding agent, and that it supersedes all previous Devstral coding models on the listed benchmarks. It uses the same reasoning_effort values as Small 4, none and high, and is dense rather than Mixture-of-Experts.
  • Large 4 (October to November 2026): moved the Large line to a 1M-token window and three reasoning efforts, low, medium and high (default). The launch post says low replaces the former "None" mode. The model card adds connectors to the API features. Mistral names Large 4 as the replacement for both Large 3 and Medium 3.5.

The practical consequence for code: a request that sets reasoning_effort to none on Medium 3.5 or Small 4 has no direct equivalent on Large 4, where low is the lowest setting. Retest prompts and parameters when moving from Medium 3.5 to Large 4. For background on effort settings, see Reasoning Models Explained.

What workloads does Mistral position each model for?

This section records the provider's own positioning wording; it is not an assessment.

ModelPositioning stated by providerRelated notes stated by provider
Mistral Large 4"Built for Coders, Trusted by Defenders" (launch post); foundation for future specialised Mistral models and for self-hosted, customised deploymentsReasoning effort guidance: high for the hardest tasks, medium for everyday tasks, low for simple tasks
Mistral Medium 3.5"Frontier-class multimodal model optimized for agentic and coding use cases" (model card)Card now directs new integrations to Large 4
Mistral Small 4"Hybrid model unifying instruct, reasoning, and coding capabilities in a single model" (model card)Launch post reports 40% lower end-to-end completion time and 3x more requests per second than Small 3 in optimised setups

Mistral's documentation cards also show provider ratings: Large 4 speed 5/5 and performance 4/5, Medium 3.5 speed 2/5 and performance 4.5/5. No rating appeared for Small 4 on the card checked. These are Mistral's own labels, not independent measurements.

Worked cost example

Using the card prices above, a workload of 10,000 requests a day, each with a 20K-token prompt and a 1K-token response, uses 200 million input and 10 million output tokens a day. On Large 4 that is $140 input plus $21 output, $161 a day. On Medium 3.5 it is $300 plus $75, $375 a day. On Small 4 it is $30 plus $6, $36 a day. Reasoning tokens are generated as output, so high effort settings raise the output count; caching on Large 4 would cut repeated input to $0.07. Tokenisers may differ between models, so measure token counts on your own prompts before comparing.

What is the context window of each model?

As of 2026-11-15. From each model's documentation card and Hugging Face card.

ModelContext windowMax outputKnowledge cutoffInputOutputReasoning controls
Mistral Large 41M tokensNot stated by Mistral AI on the pages checked (2026-11-15)Not stated by Mistral AI on the pages checked (2026-11-15)Text, imageReasoning, textreasoning_effort low, medium, high (default)
Mistral Medium 3.5256k tokensNot stated by Mistral AI on the pages checked (2026-11-15)Not stated by Mistral AI on the pages checked (2026-11-15)Text, imageReasoning, textreasoning_effort none, high
Mistral Small 4256k tokensNot stated by Mistral AI on the pages checked (2026-11-15)Not stated by Mistral AI on the pages checked (2026-11-15)Text, imageTextreasoning_effort none, high

Recommended sampling differs too. The Large 4 Hugging Face card suggests temperature 1.0 and top_p 0.95 at every effort level. The Medium 3.5 card suggests temperature 0.7 and top_p 0.95 at high, and temperature 0.0 to 0.7 with top_p unset at none. The Small 4 card suggests temperature 0.7 at high and 0.0 to 0.7 at none. Carrying one model's settings to another can change output quality. See Context Windows Explained.

Capabilities and tools compared

As of 2026-11-15. API features exactly as listed on each documentation card.

FeatureMistral Large 4Mistral Medium 3.5Mistral Small 4
Chat completionsListedListedListed
Function callingListedListedListed
Structured outputsListedListedListed
Document Q&AListedListedListed
PrefixListedListedListed
Batching (/v1/batch)ListedListedListed
Agents and conversationsListedListedListed
Other features listed on the cardConnectorsPredicted outputs; built-in tools (HTML card); connectors (Markdown card)Predicted outputs; built-in tools

The HTML and Markdown versions of the Medium 3.5 card list slightly different extras, so both are shown. Self-hosting stacks named by Mistral: vLLM (recommended for all three), SGLang and Transformers for all three, plus llama.cpp for all three. Ollama is named on the Large 4 and Medium 3.5 cards, LM Studio on the Large 4 and Small 4 cards (and as work in progress on the Medium 3.5 card), and Docker Model Runner on the Large 4 card. Mistral's published vLLM examples use tensor parallel size 8 for Large 4 and Medium 3.5 and 2 for Small 4. For Small 4 the launch post gives minimum hardware of 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200.

Where is each model available?

ModelAvailability stated by provider
Mistral Large 4Mistral Studio and Mistral API; Hugging Face weights; Microsoft Foundry (the launch post says Microsoft is the first cloud provider to offer it); NVIDIA-hosted endpoints, NIM, NeMo AutoModel and TensorRT LLM; OpenRouter; Vibe Code; Mistral Forge for customisation
Mistral Medium 3.5Mistral API (deprecated, retiring 2027-04-30); Hugging Face weights; Le Chat and the Vibe coding agent per the Hugging Face card
Mistral Small 4Mistral API and AI Studio; Hugging Face weights; NVIDIA build.nvidia.com, NIM and NeMo

Third-party platforms publish their own model IDs and lifecycle schedules, which can differ from Mistral's API. Open weights on Hugging Face stay downloadable under their licence regardless of API retirement unless Mistral states otherwise; confirm in each repository.

Model cards and system cards

ModelModel card or pageSystem cardTechnical report
Mistral Large 4Documentation card; Hugging Face cardNot stated by Mistral AI on the pages checked (2026-11-15)Hugging Face card links assets/technical_report.pdf in the repository
Mistral Medium 3.5Documentation card; Hugging Face cardNot stated by Mistral AI on the pages checked (2026-11-15)Not stated by Mistral AI on the pages checked (2026-11-15)
Mistral Small 4Documentation card; Hugging Face cardNot stated by Mistral AI on the pages checked (2026-11-15)Not stated by Mistral AI on the pages checked (2026-11-15)

Licence wording: the Large 4 launch post's model box shows "Apache 2.0" while its text and the documentation card name the Mistral Ecosystem Open Weight License; this page uses the documentation card. See Model Cards and System Cards Explained and the Open-Weight Model Licence Tracker.

Benchmarks

As of 2026-11-15. Figures are provider-reported, as stated in the text of each launch post or Hugging Face card. Suites differ between releases, so figures in different rows are not directly comparable, and no benchmark appears in the text for all three models.

ModelProvider-reported benchmarksIndependent leaderboard results
Mistral Large 4Terminal-Bench 4.0 41%; BrowseComp 84.5%; FinWorkBench (Finch) 73.3%; Cognition FrontierCode 2.0 (private) 19.7%; B3 Agent Security Benchmark 97.1% attack resistance; Artificial Analysis Coding Agent Index 48% and Cyber Index 51% as quoted by Mistral (launch post, November 2026)Not checked for this page
Mistral Medium 3.5SWE-Bench Verified 77.6%; τ³-Telecom 91.4%; LEXam-hard 31.89 (Hugging Face card, 2026); other results shown only as imagesNot checked for this page
Mistral Small 4GPQA Diamond 71.2 (Hugging Face card); AA LCR 0.72 with about 1.6K characters of output; LiveCodeBench result stated only relative to GPT-OSS 120B (launch post, 2026-03-16); other results shown only as chartsNot checked for this page

Run your own evaluation on representative tasks; provider suites, harnesses and settings vary between releases.

Deprecations and retirements

As of 2026-11-15. Source: Mistral models overview and model cards.

ModelAPI IDNotice or deprecation dateRetirement or shutdown dateStated replacement
Mistral Large 4mistral-large-4-0CurrentNone listedNone stated
Mistral Medium 3.5mistral-medium-3-52026-10-312027-04-30Mistral Large 4 (mistral-large-4-0)
Mistral Small 4mistral-small-2603CurrentNone listedNone stated

Alias risk: the Medium 3.5 Markdown card lists mistral-medium-latest and mistral-medium-3 as aliases, and mistral-large-latest points to Large 4. Floating aliases have pointed at different releases over time, so code calling them can change model without a code change. Pin the dated or versioned ID. For lifecycle planning, see the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.

Data handling and security

Mistral's help centre applies the same data rules to all three models; there is no model-specific policy on the pages checked. Zero Data Retention is available only on pay-as-you-go, only for stateless calls such as chat completions, FIM, embeddings, moderation, classification, OCR and audio, and only on request at Mistral's discretion. Agents, batch processing files, conversations, libraries, /v1/files, Vibe Work and Chat are excluded. All three models list agents, conversations and batching, so those features sit outside ZDR whichever model you choose. Self-hosted weights keep data wherever you run them. See Zero Data Retention for AI.

The launch post for Large 4 describes built-in guardrails aimed at telling defensive from malicious cyber tasks, a reported 96.6% average refusal rate on malicious cyber requests, and a separate cybersecurity variant in preview only for vetted enterprise design partners, restricted to owned or authorised systems. Mistral states no comparable variant for Medium 3.5 or Small 4 on the pages checked. Mistral API keys can call every model a workspace allows, so the key is the main operational risk; see the LLMjacking Guide and the Open-Weight Model Security Guide.

How to choose a model: factual criteria

CriterionWhat the documentation shows
Lifecycle runwayLarge 4 and Small 4 have no deprecation date; Medium 3.5 retires 2027-04-30
List price per tokenSmall 4 $0.15/$0.60; Large 4 $0.70/$2.10; Medium 3.5 $1.50/$7.50
Context above 256k tokensLarge 4 only (1M)
Reasoning off entirelyMedium 3.5 and Small 4 accept none; Large 4's lowest setting is low
Licence for self-hostingApache 2.0 (Small 4); Modified MIT with large-revenue exceptions (Medium 3.5); Ecosystem licence with a Model-as-a-Service condition (Large 4)
Self-hosting footprintSmall 4 6.5B active; Large 4 52B active; Medium 3.5 128B dense
Connectors on the APIListed on the Large 4 card and the Medium 3.5 Markdown card

For how these compare with other providers, see Open-Weight Frontier Models Compared and Frontier AI Models Compared.

Practitioner checklist

  • Inventory every service, agent and script calling mistral-medium-3-5, mistral-medium-3 or mistral-medium-latest, and migrate before 30 April 2027.
  • Pin production code to mistral-large-4-0 or mistral-small-2603 rather than a -latest alias, and record the ID in the AI inventory.
  • Retest reasoning_effort when moving from Medium 3.5 to Large 4: none is not one of Large 4's values.
  • Confirm the billed Large 4 rate in Mistral Studio, given the two official figures, and count reasoning tokens as output.
  • Read the specific licence before self-hosting: Apache 2.0, Modified MIT and the Ecosystem licence carry different conditions.
  • Check whether your use relies on agents, conversations or batch, which sit outside Zero Data Retention.
  • Store Mistral API keys in a secrets manager, separate workspaces by environment and set spend limits.

Standards and references

Related NHI Mgmt Group resources: Mistral AI Models, Mistral Large 4, Mistral Small 4, European LLMs Compared, LLM Pricing Tracker.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 15 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.