Mistral Large 4: Specs, Pricing, Benchmarks and Weights Date

TL;DR
Mistral Large 4 is a 1T-parameter multimodal MoE with 52B active, in public preview at $0.68/$2.09 per million tokens, with open weights due end of October.
Mistral Large 4 is a 1-trillion-parameter, natively multimodal mixture-of-experts model with 52 billion active parameters, in public preview on Mistral's API now, with open weights promised by the end of October 2026. It costs $0.68 input and $2.09 output per million tokens during Mistral's current 50% sale, and it is also on Vercel AI Gateway.
Last updated: October 9, 2026
The short verdict: ML4 is the strongest open-weight model built outside China on Mistral's own charts, a real step up from Mistral Large 3, and roughly six months behind the closed frontier on independent scores. It is worth evaluating today if you need EU-hosted inference, cyber-defense work that closed models refuse, or visual grounding. It is not a reason to move a coding agent off Claude or GPT-6 class models. The bigger event is the weight release at month end, and this page has a checklist for that day.
Official Sources#
| Resource | What it gives you |
|---|---|
| Mistral announcement: Introducing Mistral Large 4 | Specs, benchmark charts, the end-of-October weights promise |
| Mistral Large 4 model card | 52B active / 1.05T total, 1.6B vision encoder, context, prices, supported endpoints |
| Mistral API pricing | List and sale prices next to Mistral Large 3, Medium 3.5 and Mistral-hosted GLM 5.3 |
| Mistral API reference: chat completions | Request format and the reasoning_effort parameter |
| Mistral reasoning docs | How reasoning_effort works and what reasoning responses return |
| Vercel changelog: ML4 on AI Gateway | Same-day gateway access as mistral/mistral-large-4 |
| Artificial Analysis model page | Independent intelligence index, speed, verbosity and cost per task |
Specs at a glance#
| Spec | Mistral Large 4 (preview) |
|---|---|
| Architecture | Granular mixture-of-experts, hybrid instruct and reasoning |
| Parameters | 1.05T total, 52B active, plus a 1.6B vision encoder (model card) |
| Input / output | Text and image in, text out |
| Context | 1M listed on Mistral's model card; 524K listed by AI Gateway and Artificial Analysis |
| Max output | 262K tokens on AI Gateway |
| Reasoning | Adjustable via reasoning_effort; none and high reported working on ML4 |
| API model ID | mistral-large-4 (Mistral), mistral/mistral-large-4 (AI Gateway) |
| Training | From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters |
| Languages | Training data spans 160+ languages, including every official EU language |
| Weights | Promised "by the end of the month" (October 2026), with architecture details and more benchmarks |
| Status | Public preview; the RL run is still in progress, per the announcement |
Two numbers moved after launch, so check which one a source uses. Launch-day coverage, including Simon Willison's notes, Plotly's benchmark write-up and the first version of this post, reported 49B active parameters. Mistral's model card now says 52B active and 1.05T total, and the announcement says 52B active in a 1 trillion-parameter model. We use Mistral's current figure. The context window is the other mismatch: the model card says 1M, while both AI Gateway and Artificial Analysis list 524K for the served preview. Plan around 524K until Mistral's own endpoint proves otherwise.
Benchmarks (vendor-reported unless marked)#
Every number in this first table comes from Mistral's announcement. Several are run by third parties (Artificial Analysis, vals.ai, Surge AI) but chosen and published by Mistral, so treat them as vendor-selected.
| Area | Benchmark | ML4 score | Mistral's comparison |
|---|---|---|---|
| Coding | DeepSWE v1.1 | 61.7% | - |
| Coding | SWE-Atlas-QnA | 59.4% | - |
| Coding | Terminal-Bench 4 | 28.3% | - |
| Coding | Coding Agent Index | 49.8% | Ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max |
| Coding (human eval) | Surge AI blind rating, 1-5 | 3.74 | Claude Opus 5 4.22, GLM-5.3 3.60, Kimi K3 3.59, GLM-5.2 3.40 |
| Agents | AutomationBench (657 business workflows) | 59.9% | Ahead of Kimi K3, MiMo-V2.6-Pro, DeepSeek V4 Pro |
| Knowledge work | AA-Briefcase | 1,393 Elo | Ahead of DeepSeek V4 Pro |
| Cyber | Vulnerability reproduce-and-patch test in the AA Cyber Index | 82% | Highest of any model; Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse |
| Cyber | Cybench (40 CTF-style challenges) | 93% | - |
| Vision | Dense 200 (visual grounding) | 42% | GPT-6 Astra 41% |
| Safety | Lakera B3 attack resistance | 93.3% | Highest Mistral saw among competitors |
| Safety | KORA Benchmark | 1.691 of 2 | Highest Mistral measured among open models |
Read the coding rows carefully. The Surge human evaluation is the most useful line in the whole announcement because it includes a closed model: ML4 beats the Chinese open-weight models it was compared against, and trails Claude Opus 5 by almost half a point on a five-point scale. That matches the independent picture below. The cyber headline needs the same care. Part of ML4's lead on that test comes from closed models refusing the task, not failing it, which is exactly Mistral's pitch to defenders and exactly what one Hacker News commenter flagged as an offensive-use question.
Independent numbers so far, from Artificial Analysis (measured on Mistral's API, preview build):
| Metric | ML4 preview | Artificial Analysis comparison |
|---|---|---|
| Intelligence Index v4.3.2 | 38 | Median 26 for its price class |
| Output speed | 116.1 tokens/s | Median 74.6 |
| Time to first token | 1.46 s | Median 3.84 s |
| Output tokens to run the index | 200M | Median 81M (very verbose) |
| Cost per index task | $1.13 | - |
Simon Willison puts the 38 in context: it sits just behind DeepSeek 4.1 Flash, a 552B model, and is a big jump from Mistral Large 3, which scored 9. His summary is that Mistral is "back to being maybe about 6 months behind the frontier." He also found the reasoning setting made little visible difference: high used 2,717 output tokens on his test prompt versus 3,275 for none.
Pricing#
Mistral's pricing page lists ML4 with a struck-through list price and a sale price. Mistral does not state an end date for the sale, so budget at list price.
| Model (Mistral API, standard tier) | Input /1M | Cached input /1M | Output /1M |
|---|---|---|---|
| Mistral Large 4 (sale) | $0.68 | $0.07 | $2.09 |
| Mistral Large 4 (list) | $1.36 | $0.14 | $4.18 |
| Mistral Large 3 | $0.50 | $0.05 | $1.50 |
| Mistral Medium 3.5 | $1.50 | $0.15 | $7.50 |
| Z.ai GLM 5.3 (hosted by Mistral) | $1.40 | $0.14 | $4.40 |
The same page also has batch and priority tiers. AI Gateway passes through the sale price: its model page lists Mistral as the only provider at $0.68 input, $0.07 cache read and $2.09 output.
Worked cost example#
Take one agent session that reads 2M input tokens (1.6M of them cache hits, 0.4M uncached) and writes 100K output tokens:
| Route | Calculation | Cost |
|---|---|---|
| ML4, sale | 0.4 x $0.68 + 1.6 x $0.07 + 0.1 x $2.09 | $0.59 |
| ML4, list | 0.4 x $1.36 + 1.6 x $0.14 + 0.1 x $4.18 | $1.19 |
| GLM 5.3 on Mistral | 0.4 x $1.40 + 1.6 x $0.14 + 0.1 x $4.40 | $1.22 |
| Kimi K3, Moonshot first-party | 0.4 x $3.00 + 1.6 x $0.30 + 0.1 x $15.00 | $3.18 |
Kimi K3's rates are Moonshot's listing on AI Gateway. The caveat is verbosity. Artificial Analysis measured ML4 producing about 2.5 times the median output tokens on its index, so the real output line can be bigger than the 100K assumed here. Run your own task and read the usage field before trusting any per-token comparison.
How to call it today#
There are two routes live: Mistral's own API and Vercel AI Gateway. The Vercel snippet is verbatim from its changelog; the Mistral curl follows the API reference's example with the model ID, prompt and reasoning_effort changed. We have not run either against ML4.
Mistral API, adapted from the chat completions reference and the reasoning docs:
curl https://api.mistral.ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $MISTRAL_API_KEY" \
-d '{
"model": "mistral-large-4",
"reasoning_effort": "high",
"messages": [
{"role": "user", "content": "Find the off-by-one bug in this function and explain the fix."}
]
}'
With reasoning_effort: "high" the response includes a thinking chunk before the answer; with none it returns a plain string. Mistral's reasoning docs document none, low, high and max across models and list ML4 as supporting adjustable reasoning without naming its values. Simon Willison reports ML4 accepts only none and high, so test low and max before relying on them.
Vercel AI Gateway, verbatim from the changelog:
import { streamText } from 'ai';
const result = streamText({
model: 'mistral/mistral-large-4',
prompt: 'Fix the failing test and verify the change.',
});
The gateway also exposes ML4 through its OpenAI-compatible Chat Completions, Responses and Anthropic Messages APIs. For a coding agent, Vercel's documented path is npx vercel ai-gateway setup, then selecting mistral/mistral-large-4 in the agent's model settings. If your agent is Codex, the provider wiring is the same model_providers pattern we cover in using custom model providers in Codex: a base URL, an env key and the model ID.
Who gets squeezed when a 1T open-weight model lands#
GLM-5.3 is the most directly exposed. Mistral already resells GLM 5.3 on its own platform at $1.40/$4.40. ML4's list price sits just under that, and the sale price is half. Mistral is using its own price sheet to tell customers who adopted a Chinese open model through Mistral that there is now a home-grown option at the same or lower cost. One HN commenter describes exactly that path: Mistral Medium 3.5, then Mistral-hosted GLM-5.3, now wondering whether to switch again. GLM-5.3 still has the stronger price floor elsewhere, through Flash and cheap third-party hosts, which we track in where to run GLM-5.3 free and cheap.
Kimi K3 is exposed on price, not capability. K3 lists at $3/$15 first-party, roughly 2.2x ML4's list input price and 3.6x its list output price. On Mistral's own human eval the two are close on coding. For buyers who want a large open model with European provenance, that gap is hard to ignore.
The closed mid-tier is exposed on refusals and residency. Mistral's cyber pitch is aimed straight at teams whose vulnerability research gets blocked by closed-model refusals, and its European deployment that Mistral says it runs "independently of other digital service providers" is aimed at procurement teams that cannot send data to US clouds.
Who is not squeezed: the cheapest models and the frontier. DeepSeek 4.1 Flash scores slightly higher on Artificial Analysis, per Simon Willison, and one HN commenter noted it remains the cheapest model in these comparisons by a long way. Claude Opus 5 still leads ML4 clearly on Mistral's own coding human eval.
The second-order effect is what happens to Mistral's price after the weights ship. Look at Kimi K3: once its weights were public, AI Gateway alone now lists 17 providers for it, with input prices from $3 down to $0.45 and output from $15 down to $11.40 per million tokens. GLM-5.3's cheapest third-party input rate fell to $0.12 in our October 1 check. When ML4's weights land, independent hosts will compete on the same model, and Mistral's own API will have to win on reliability, EU residency and support rather than on being the only door. Our read is that the 50% launch sale is partly about anchoring the reference price at $0.68/$2.09 before that competition arrives.
When the weights drop: a checklist#
Mistral says it will release weights "by the end of the month" along with architecture details, more benchmarks and its post-training method. Guillaume Lample, Mistral's co-founder, posted (as quoted on the Hacker News thread) that the RL run is still going and that they will "release a final version before the end of the month along with the weights." So the downloadable model may not be the same build as today's preview. On release day, check these in order:
- The license. Apache 2.0, a Mistral research license, or something custom decides whether you can use it commercially, fine-tune it, and redistribute derivatives. Artificial Analysis currently labels the preview "Proprietary" because no weights are public.
- Which build shipped. Compare the release notes against the preview: same 52B active / 1.05T total, same vision encoder, and whether the "final version" moved benchmarks.
- Precision and size. At 1.05T parameters, the weights alone are roughly 1 TB at 8 bits per parameter and roughly 525 GB at 4 bits, before KV cache. This is multi-GPU server hardware, not a workstation. For local work, a model like Gemma 4 is still the realistic choice.
- Serving runtimes. Check whether vLLM, SGLang and llama.cpp list day-one support for the architecture, and which quantizations Mistral publishes itself.
- Context in practice. Confirm whether the open checkpoint supports the 1M context on the model card or the 524K the preview serves.
- Third-party hosts and prices. Watch how quickly hosts appear on AI Gateway and other routers, and whether they undercut $0.68/$2.09.
- Independent evals on the final build. Re-run your own repo tasks and wait for Artificial Analysis to re-score it as open weights.
We will update this page with each answer when the weights land.
Decision guide#
| If you need | Use ML4? | Why |
|---|---|---|
| EU-hosted, EU-trained model for regulated work | Yes, evaluate now | Trained and served on Mistral's own European infrastructure, with an EU deployment Mistral operates end to end |
| Defensive security research blocked by refusals | Yes, evaluate now | 82% on the reproduce-and-patch test where Mistral says closed models refuse |
| Visual grounding, charts, documents, engineering drawings | Yes, evaluate now | Mistral's strongest relative claims; Dense 200 edges GPT-6 Astra |
| Best coding-agent quality | No | Claude Opus 5 leads ML4 by about half a point on Mistral's own coding human eval |
| Lowest cost per token | No | DeepSeek 4.1 Flash scores slightly higher on Artificial Analysis for less; GLM-5.3-Flash lists at $0.15/$0.50 |
| Self-hosting | Wait | No weights or license yet; check the list above at month end |
If you already route models, add ML4 behind a feature flag and run five tasks from your own backlog: a long-context repo question, a small bug fix with tests, a terminal-heavy setup task, one image or PDF task, and a failure case from your current model. Router setups make this cheap to try, which is the argument in model routers and the optionality advantage.
What people are actually saying#
The main Hacker News thread passed 2,000 points and 1,200 comments by October 9; a second submission reached 522 points before its comments were merged into the first.
- The positive read is about vision and cyber. One commenter called the vision and cyber numbers good enough for a daily driver for people who want to avoid the alternatives. Others framed it as a sovereignty milestone: trained in the EU, served in the EU.
- The most useful practitioner report came from Plotly. Its team ran ML4 through a private 43-question data analytics exam: 32 of 43 passed (74%) for $1.94 a run, up from 25 of 43 (58%) for $20.51 with Mistral Medium 3.5. Its misses were judgment calls, such as reporting a 7-day total as an average and missing duplicated rows hidden inside a correct-looking count. The author's HN comment adds that it is not yet on the cost-quality frontier.
- The sharpest disagreement is price-for-quality. A commenter citing the Vals Index put ML4 at 48.05% versus GLM-5.3 at 53.51%, at $13.78 versus $7.25 per test. Another called it comparable to GPT-6 Luna and DeepSeek 4.1 Flash but pricier than both, and a third compared the enthusiasm to cheering for the last kid across the finish line.
- Benchmark selection drew questions. One reply asked why Mistral's chart puts GLM 5.3 at 61 on DeepSWE 1.1 when other listings show 69. Another objected to "open weight but closed science" until Mistral publishes its recipe.
- A hands-on coding read sits in the middle. A developer with a custom-language benchmark found ML4 roughly level with Kimi K3 and Claude Sonnet 5.5 on it, but much slower than Sonnet 5.5 and more expensive than both.
The counter-case to this whole page: if the weights slip, the license is restrictive, or the final build underperforms the preview, ML4 is a decent mid-priced API in a crowded tier rather than an open-weight event.
FAQ#
What is Mistral Large 4?#
Mistral Large 4 (ML4, nicknamed "le Chonk") is Mistral's largest model, announced October 6, 2026. It is a natively multimodal mixture-of-experts model with 1.05T total parameters, 52B active, and a 1.6B vision encoder, released as a public preview with open weights promised by the end of October 2026.
When will the Mistral Large 4 weights be released?#
Mistral says by the end of October 2026, after red-teaming with security partners and state authorities. Co-founder Guillaume Lample said a final version will ship with the weights, so the downloadable checkpoint may differ from today's preview.
How much does Mistral Large 4 cost?#
On Mistral's API it lists at $1.36 input, $0.14 cached input and $4.18 output per million tokens, currently on sale at $0.68, $0.07 and $2.09. Vercel AI Gateway passes through the same sale price. Mistral has not published an end date for the sale.
Is Mistral Large 4 49B or 52B active parameters?#
Mistral's model card says 52B active and 1.05T total, and the announcement says 52B active in a 1 trillion-parameter model. Launch-day coverage, including Simon Willison, Plotly and the first version of this post, reported 49B. Use 52B.
How good is Mistral Large 4 at coding?#
Mistral reports 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and a 3.74 blind human rating against Claude Opus 5's 4.22. Independent testing so far puts it near DeepSeek 4.1 Flash and behind the closed frontier, so evaluate it on your own repository tasks before switching.
Can I use Mistral Large 4 through Vercel AI Gateway?#
Yes. It has been on AI Gateway since October 6 as mistral/mistral-large-4, usable through the AI SDK, OpenAI-compatible Chat Completions, the Responses API and the Anthropic Messages API, with Mistral as the serving provider.
Continue Reading#
- Where to Run GLM-5.3 Free and Cheap - the open model ML4 is priced against
- How to Use GLM 5.2 and Other Custom Model Providers in Codex - wire ML4 or any OpenAI-compatible endpoint into Codex
- Gemma 4: The Open Model Guide for Developers - the open-weight option you can actually run locally
- Where to Access Kimi K3: Every Provider and Price Compared - every K3 provider compared
- Model Routers and the Optionality Advantage - why adding ML4 to a router is cheap
- Tencent Hy4 Preview: 770B Open MoE, Agentic Benchmarks, and Where It Fits in 2026 - Tencent's 770B Apache 2.0 open MoE with 49B active, 1M context, agentic benchmarks and OpenRouter pricing
Sources#
| Source | URL |
|---|---|
| Mistral: Introducing Mistral Large 4, fetched October 9, 2026 | https://mistral.ai/news/mistral-large-4/ |
| Mistral docs: Mistral Large 4 model card, fetched October 9, 2026 | https://docs.mistral.ai/models/mistral-large-4-0 |
| Mistral docs: API pricing, fetched October 9, 2026 | https://docs.mistral.ai/inference/pricing |
| Mistral docs: chat completions API reference | https://docs.mistral.ai/api/endpoint/chat |
| Mistral docs: reasoning | https://docs.mistral.ai/studio/conversations/reasoning |
| Vercel changelog: Mistral Large 4 on AI Gateway, October 6, 2026 | https://vercel.com/changelog/mistral-large-4-now-available-on-ai-gateway |
| Vercel AI Gateway: Mistral Large 4 model page, fetched October 9, 2026 | https://vercel.com/ai-gateway/models/mistral-large-4 |
| Vercel AI Gateway: Kimi K3 model page, fetched October 9, 2026 | https://vercel.com/ai-gateway/models/kimi-k3 |
| Artificial Analysis: Mistral Large 4 Preview, fetched October 9, 2026 | https://artificialanalysis.ai/models/mistral-large-4 |
| Artificial Analysis: Mistral Large 3 | https://artificialanalysis.ai/models/mistral-large-3 |
| Simon Willison: Introducing Mistral Large 4: Le chonk, October 6, 2026 | https://simonwillison.net/2026/Oct/6/le-chonk/ |
| Plotly: Mistral Large 4 data analytics benchmark results | https://plotly.com/blog/mistral-large-4-plotly-data-analytics-bench/ |
| Guillaume Lample on X, as linked in the HN thread | https://x.com/GuillaumeLample/status/2107461898127954001 |
| Hacker News: Mistral Large 4 (main thread), checked October 9, 2026 | https://news.ycombinator.com/item?id=49977979 |
| Hacker News: Mistral Large 4: "Le Chonk" (merged thread) | https://news.ycombinator.com/item?id=49978116 |
Get the next comparison like this in your inbox
One email a week on Mistral and the rest of the AI dev stack. Free.
Read next on local and open-weight models
GLM-5.3 Free and Cheap: Every Route and Price (October 2026)
GLM-5.3 has no free hosted tier. Cheapest on October 9: Wafer via OpenRouter at $0.049/$3.39, Flash from $0.075/$0.25. Z.ai's promos ended October 7.
13 min readCodex Custom Model Providers: GLM-5.3, Mistral and Ollama
Codex only speaks the Responses API now. Tested config.toml blocks for GLM-5.3, Mistral Large 4 through a gateway, local Ollama and an internal proxy.
14 min readGemma 4: The Open Model Guide for Developers
Gemma 4 ships byte-for-byte open weights from Google DeepMind. How developers deploy it locally, fine-tune it, and ship agents on top of it.
11 min readNew here? Start with
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.








