MistralLarge4.org

Mistral Large 4 Timeline

Mistral Large 4 was released as a public preview on the Mistral API on October 6, 2026, appeared on OpenRouter and Vercel AI Gateway the same day, and Mistral AI says its open weights will follow by the end of October 2026.

Updated 2026-10-06

Mistral Large 4 at a glance

Developer
Mistral AI (Paris, France)
Released
October 6, 2026 (public preview)
Parameters
1.05T total, 49B active per token (Mixture of Experts)
Context window
524,288 tokens via the API
Max output
262,144 tokens
Input / output
Text and images in, text out
API price (sale)
$0.68 input / $2.09 output per 1M tokens (list $1.36 / $4.18)
API model name
mistral-large-4
Reasoning
reasoning_effort: "high" or "none"
Open weights
Scheduled for the end of October 2026

Sources: Mistral AI announcement and pricing page, OpenRouter model listing, Hugging Face model page. Checked Oct 6, 2026.

Current status as of October 6, 2026

Mistral Large 4 is usable today through the Mistral API, OpenRouter, Vercel AI Gateway and the chat on our homepage; the weights, license and general-availability date are the open items.

ItemStatus on Oct 6, 2026
AccessPublic preview (version v26.10) on the Mistral API, OpenRouter and Vercel AI Gateway
API priceList $1.36 / $4.18 per 1M tokens; temporary 50% sale at $0.68 / $2.09
Context served524,288 tokens (Mistral docs list 1M); max output 262,144
Open weightsNot yet published; Hugging Face repo listed as an upcoming release
LicenseNot announced
Inference enginesNo vLLM, SGLang or llama.cpp support announced
Vibe (formerly Le Chat)Mistral has not announced Large 4 in Vibe
Azure, Bedrock, Vertex, watsonxNot announced
Chat on mistrallarge4.orgLive: 3 free messages a day, 15 credits a day with a free account
Source: Mistral launch post, Mistral Docs, Hugging Face, OpenRouter, Vercel, GitHub issue search, checked Oct 6, 2026.

Release timeline with sources

Every row below is a dated event from a public source. Planned dates are marked as expected.

DateEventSource
Dec 2, 2025Mistral Large 3 (675B total, 41B active, Apache 2.0) releasedVals AI, OpenRouter
Aug 12, 2026Le Chat renamed Vibe, with Work, Code and Chat modesMistral Help Center
Sep 2026Mistral raises a €3B Series D led by Samsung at about a €21B valuation; the Large 4 post calls the model the first milestone it fundsMistral launch post, Constellation Research
Oct 6, 2026Mistral Large 4 public preview on the Mistral API, model IDs mistral-large-4 and mistral-large-4-0Mistral launch post, Mistral Docs
Oct 6, 2026Pricing published: $1.36 input, $4.18 output, $0.14 cached per 1M tokens, with a temporary 50% saleMistral Docs pricing page
Oct 6, 2026Available on OpenRouter (mistralai/mistral-large-4-0) and Vercel AI Gateway (mistral/mistral-large-4)OpenRouter, Vercel
Oct 6, 2026Mistral publishes launch benchmarks, including 61.7% on DeepSWE v1.1 and 59.9% on AutomationBenchMistral launch post
By Oct 6, 2026Artificial Analysis lists an Intelligence Index score of 38 and 116.1 output tokens per secondArtificial Analysis
By Oct 6, 2026Vals AI lists a Vals Index of 48.05% (#32 of 44)Vals AI
By Oct 6, 2026Hacker News launch thread passes about 1,045 points and 689 commentsHacker News
Oct 27, 2026 (expected)Weights date reported by The Next Web, Forkast and a models.dev pull requestPress reports
Oct 31, 2026 (expected)Weights ETA shown on the Hugging Face repo mistralai/Mistral-Large-4.0-1T05-A52BHugging Face
Source: as listed per row, checked Oct 6, 2026.

What Mistral announced on launch day

Mistral introduced Large 4, nicknamed le Chonk, as a 1.05T-parameter granular Mixture-of-Experts model with 49B active parameters per token, text and image input, and one model that handles both instant answers and reasoning.

  • Training: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters (TechCrunch and The Next Web round this to about 4,000).
  • Reinforcement learning at about 33B tokens a day per training run, around 16B of them trainable completion tokens.
  • Training data in more than 160 languages, including every official EU language.
  • Cyber leaders, vetted partners and state authorities get a build with reduced moderation for red-teaming before the weights release.
  • API features at launch: function calling, structured outputs, document QnA, prefix, batching, agents and conversations, and built-in tools.

First independent results and reactions

Within a day of launch, two independent evaluators had scored the model. Artificial Analysis gave it 38 on Intelligence Index v4.3.2 (#64 of 225 models, 8th among open-weight models), measured 1.46 s to first token and noted it used 200M output tokens to run the index against a median of 81M. Vals AI ranked it #6 of 75 on Harvey's Legal Agent Benchmark (15.83%) and measured 22.73% on Terminal-Bench 4.0, below the 28.3% in Mistral's own chart.

On Hacker News the launch thread was mixed: many welcomed a competitive European lab and the vision and cyber results, while others judged the scores mid-pack for a 1T model at this price.

Dates to watch next

The next fixed milestone is the open-weights release at the end of October 2026; Mistral has not given dates for the other open items yet.

  • Weights: Mistral says by the end of October; the Hugging Face ETA is October 31 and press reports say October 27. The 1.05T-parameter FP8 download would be about 1.05 TB (1.05e12 × 1 byte); see /run-locally.
  • License: Mistral has not announced it; Large 3 used Apache 2.0. See /huggingface.
  • Sale price: Mistral labels $0.68 / $2.09 a temporary sale and has not announced an end date. See /api-pricing.
  • General availability: the model is labelled Public Preview v26.10 with no GA date announced.
  • Engine support: vLLM, SGLang and llama.cpp support usually follows the weights; none is announced yet.

Frequently asked questions

More about Mistral Large 4

API PricingMistral Large 4 API costs $1.36/$4.18 per 1M tokens, now 50% off at $0.68/$2.09. Estimate your bill and compare with DeepSeek V4, GPT-6 and Claude.BenchmarksMistral Large 4 benchmark scores: 61.7% DeepSWE, 59.9% AutomationBench, AA Index 38, Vals Index 48.05%. Mistral and independent results side by side.API GuideUse the Mistral Large 4 API with model ID mistral-large-4: cURL, Python and TypeScript code, reasoning_effort, image input, tools, JSON and OpenRouter.Run LocallyMistral Large 4 weights need about 1.05 TB at FP8 and 525 GB at 4-bit. GPU setups that fit, KV cache math and what to do before the weights ship.Weights & LicenseMistral Large 4 open weights are due by the end of October 2026 at mistralai/Mistral-Large-4.0-1T05-A52B. Status, release date, license and download size.Use It FreeYes: chat with Mistral Large 4 free on our homepage, 3 messages a day with no sign-up or 15 credits a day with a free account. Every other option compared.vs DeepSeek V4Mistral Large 4 vs DeepSeek V4 Pro: near-identical index scores, image input vs text-only, 524K vs 1M context, and what each costs at peak and off-peak.vs Large 3Mistral Large 4 vs Large 3: 1.05T vs 675B parameters, built-in reasoning, 2x the context, 160+ languages, and what the higher price means for your bill.FAQShort, sourced answers about Mistral Large 4: parameters, context window, API price, weights date, license, image input, hardware and free access.