by Mistral AI
Mistral AI's largest model to date, a reasoning-capable multimodal MoE for coding, agents and cybersecurity, with public weights under a modified MIT license.
Mistral Large 4 left preview in early November 2026 with downloadable weights under a modified MIT license that only restricts resale as a hosted service. It scores 26 on Artificial Analysis Intelligence Index v5.0.1, above the 23 median for open-weight models its size, and serves a 1M-token context.
Checked 9 Nov against Mistral AI's own pages.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.021 | $0.0021 | $0.023 |
| Support reply | $0.0014 | $0.0006 | $0.0020 |
| One coding agent run | $0.140 | $0.042 | $0.182 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Mistral lists the API at $0.70 per 1M input tokens, $0.07 per 1M cached input tokens and $2.10 per 1M output tokens, read on its pricing page on 9 November 2026. During the October preview the page showed a 50% sale price of $0.68 input and $2.09 output against crossed-out list prices of $1.36 and $4.18; that sale label no longer appears. Mistral Large 3 is listed at $0.50 input and $1.50 output per 1M tokens. Batch and Priority rates sit on separate tabs of Mistral's pricing page and are not recorded here. The open weights can also be self-hosted with no per-token fee.
Mistral Large 4 is Mistral AI's 1.05 trillion parameter open-weight Mixture-of-Experts model, generally available since November 2026 after an October preview. About 49 billion parameters fire per token. It takes text and images, serves a 1M-token context, and costs $0.70 per million input tokens on Mistral's API, read 9 November 2026.
Reads
Writes
Granular Mixture-of-Experts
Sparse routing picks a handful of experts per token, so roughly 5% of the weights do the work on each step.
1M-token context
Prompts up to 1,048,576 tokens and replies up to 256K tokens, per the model card and OpenRouter.
Three reasoning levels
Low, medium and high reasoning effort per request, with high as the default.
Agent and tool support
Function calling, strict JSON schemas, document question answering, batching, and Agents and Conversations endpoints with built-in tools.
Open weights with NVFP4
Weights on Hugging Face plus an official NVFP4 checkpoint for Blackwell GPUs, served through vLLM or SGLang.
Cybench challenges solved (vendor-reported)
Source: Mistral Large 4 launch post, 6 October 2026Priced around the middle of the 106 GA models with a published price (rank 45), rank 24 of 59 on output speed as cited from Artificial Analysis, and one of 41 that document a zero-data-retention option.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $1.05/M | #45 / 106 | peer median $1.73/M |
| Output speedtokens/s | 112 tok/s | #24 / 59 | peer median 94 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Mistral AI | 1M | $0.70 / $2.10 | text, image → text | |
| Anthropic | 1M | $4 / $20 | text, image, pdf → text | |
| 1M | $2 / $10 | text, image, video, audio → text | ||
| Tencent | 1M | $0.07 / $0.281 | text, image → text |
Training data
Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, with a significant share of multilingual data across more than 160 languages, then post-trained with supervised fine-tuning and large-scale reinforcement learning. Mistral has not published a knowledge cutoff.
| Name | Endpoint | Auth |
|---|---|---|
| Mistral Studio API (model id mistral-large-4) | https://docs.mistral.ai/models/mistral-large-4-0 | API key |
| OpenRouter (Mistral as sole provider) | https://openrouter.ai/mistralai/mistral-large-4-0 | OpenRouter API key |
| Hugging Face weights (self-host with vLLM or SGLang) | https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B | None (public weights) |
Match reasoning effort to the task
Use reasoning_effort "low" for chat replies and extraction, "medium" for day-to-day work and "high" for agent runs.
Mistral's model card recommends this split and makes high the default, which spends the most tokens.
Cap output length explicitly
Set max_tokens well below the 256K ceiling for chat and tool-call turns.
Artificial Analysis rates the model somewhat verbose, so uncapped completions raise per-task cost.
Two release dates in circulation
Artificial Analysis dates the release 29 October 2026, while Mistral's model card marks version 26.10 GA on 5 November 2026; the hosted preview opened on 6 October.
Workaround: Cite the date for the event you mean: preview, weights or GA.
High reasoning is the default
Reasoning effort defaults to high, the most token-hungry setting, and Artificial Analysis counted 107 million output tokens on its index.
Workaround: Set reasoning_effort to low or medium for routine calls and cap max_tokens.
Open weights, restricted resale
The weights are downloadable under a modified MIT license, but a Model as a Service business needs a separate commercial license.
Workaround: Check whether your product gives third parties meaningful control over the model before you ship it as an API.
Parameter count rounds differently
Mistral's docs list 1.05T total and 52B active, while the Hugging Face card says about 1T total and 49B active excluding embeddings and output layers.
Workaround: Size hardware from the checkpoint files, not the headline number.
mistral-large-4-0 (v26.10, GA) · 05 Nov 2026
Generally available with public weights on Hugging Face under the Mistral Ecosystem Open-Weight License, 1M context, NVFP4 checkpoint.
mistral-large-4-0 (v26.10, preview) · 06 Oct 2026
Hosted public preview: 1.05T total and 52B active parameters, image input, sale-priced API.
mistral-large-3 (v25.12) · 02 Dec 2025
Predecessor: 675B total and 41B active parameters, listed by Mistral under Apache 2.0.
Mistral lists $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens on its pricing page, read on 9 November 2026. The October preview carried a 50% "sale" label against crossed-out prices twice as high; that label is gone and the listed rate is now the only one shown. OpenRouter passes the same rates through with Mistral as the sole provider.
On Mistral's own charts it leads DeepSeek V4 Pro 0813 on its Coding Agent Index and on AutomationBench, but those are vendor results. On the independent Artificial Analysis aggregate, Mistral's own model card concedes that the leading Chinese open-weight models score higher. Pick Large 4 for image input, cyber work and EU hosting; pick DeepSeek for general reasoning per dollar under its MIT weights.
Yes, in most cases. The Mistral Ecosystem Open-Weight License is a modified MIT license that allows commercial use, fine-tuning and redistribution. The exception is a Model as a Service business that sells third parties API or hosted access to the model: it needs a separate commercial license from Mistral. Internal use and products that embed the model in specific features are exempt.
Mistral's help center says Studio free mode may use inputs and outputs for training with an opt-out, pay-as-you-go customers can opt out at any time, and Vibe Enterprise is opted out by default. Thumbs-up or thumbs-down feedback lets Mistral use that exchange. A Zero Data Retention option is documented, and self-hosting the open weights keeps data off Mistral's servers entirely.
It fits security teams, EU organizations that want a European-run or self-hosted model, and agent builders who need image input with tool calls. Pass if you only want the most capable model per dollar, if you cannot staff an 8-GPU deployment and do not want the API, or if you resell inference and do not want to negotiate a commercial license.
Mistral Large 4 is the largest model in Mistral AI's Large line and, since early November 2026, a generally available open-weight release rather than a preview. Mistral first opened it on 6 October 2026 as a hosted preview nicknamed "le Chonk". Artificial Analysis dates the weights-backed release 29 October 2026, and Mistral's model card now marks version 26.10 as GA, dated 5 November 2026. It is a granular Mixture-of-Experts model with 1.05 trillion total parameters per the model card (the Hugging Face card rounds this to about 1 trillion), 49 billion active per token or 52 billion counting embeddings and output layers, 128 routed experts with 4 active plus 1 shared, and a 1.6 billion parameter vision encoder. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. It follows Mistral Large 3, the 675B total, 41B active model from December 2025, and is built by Mistral AI.
The weights are public on Hugging Face as Mistral-Large-4-1T-A52B under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. It allows commercial and non-commercial use, modification, fine-tuning and redistribution, with one condition: a company that runs a Model as a Service business, meaning it gives third parties API or hosted access with meaningful control over inputs, parameters or training data, must sign a separate commercial license with Mistral first. Internal use, end-user products that embed the model inside specific features, and use through Mistral's own products are exempt. That makes it more restrictive than Large 3's Apache 2.0 for hosting providers, and close to MIT for everyone else. Artificial Analysis files it under open weights with commercial use restricted.
Mistral publishes most results as charts, so every number in this paragraph is vendor-reported unless it names another source. On agentic coding Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, and places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max on its own Coding Agent Index. Its Hugging Face card cites a 48% score on the Artificial Analysis Coding Agent Index, which it calls effectively tied with GLM-5.3 while using 47% fewer tokens per task. In cybersecurity it reports 93% on Cybench and 82% on CyberGym-E2E, a test that asks a model to reproduce a real vulnerability and then patch it. A blind Surge AI rating of coding output put it second of five at 3.74 out of 5, behind Claude Opus 5 and ahead of Kimi K3.
The independent view is more modest. Artificial Analysis now scores it 26 on Intelligence Index v5.0.1, a new nine-evaluation version of the index, against a median of 23 for open-weight models of its size, so scores from the older v4 index (where the preview scored 38) do not compare. It measured 112 output tokens per second and 107 million output tokens to finish the index, against a median of 99.3 million, and it describes the model as somewhat verbose. Running the index cost $2.06 per task at Mistral's rates. The model leaderboard sets those scores next to current prices.
The context window is 1M tokens on the model card, on Artificial Analysis and on OpenRouter, which also lists a 256K output cap, so the 524K figure the preview endpoint served in October no longer applies. Input is text and images and output is text. The API supports function calling, structured outputs with strict JSON schemas, document question answering, prefix completion, batching, and the Agents and Conversations endpoints with built-in tools. Reasoning effort has three settings, low, medium and high, with high as the default. Mistral says the training data covers more than 160 languages, including every official EU language.
Self-hosting is now possible but heavy. Mistral recommends vLLM and ships an official Docker image, its examples assume an 8-GPU node with tensor parallelism, and a quantized NVFP4 checkpoint targets Blackwell GPUs with a smaller memory footprint. Only about 5% of the parameters fire per token, which keeps compute down, but the full weights still have to sit in GPU memory, which rules out a single workstation. The local AI guide covers what smaller setups can run instead.
On safety, Mistral reports 93.3% attack resistance on Lakera's B3 agent security benchmark and 1.691 of a possible 2 on the KORA benchmark, and says its refusal rate on harmful cyber prompts is higher than any open-weight model it compared. It also says Claude Opus 5.5 and GPT-6 Astra score near zero on one cyber test because they refuse it, which is Mistral's claim, not an independent finding.
Who should not pick it: a team that only wants the cheapest capable API gets more general-index score per dollar from smaller models, and the leading Chinese open-weight models score higher on the Artificial Analysis aggregate, which Mistral's own card concedes. A hosting provider that plans to resell inference needs Mistral's commercial license before launch, and a team that wants plain Apache 2.0 terms can still run Mistral Large 3. Hokai's verdict on what Large 4 is for walks through the jobs where it does fit: security work, EU-sovereign deployments and self-hosted agents.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Mistral AI's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.