Mistral Large 4 is the biggest model the French company has built. Mistral released it as a public preview on Tuesday 6 October 2026: a mixture-of-experts model with 1 trillion parameters, of which 49 billion are active for each token, trained from scratch in Mistral’s own European data centres. Mistral’s own nickname for it is “le Chonk”. The preview runs on Mistral’s API today, and Mistral says the weights will follow “by the end of the month”, which several outlets reported as 27 October.
The launch is Mistral’s clearest statement yet of its sovereign AI pitch. In a statement reported by AFP, Mistral said enterprises “across financial services, manufacturing and the public sector can govern the intelligence, data, compute and operations” without sending sensitive knowledge outside. Mistral calls it the first milestone funded by the €3 billion Series D that valued Mistral at more than €21 billion in September.
Below we set out what Mistral Large 4 is, where the published figures disagree, why Mistral leans so hard on cyber defence, how its benchmark claims compare with an independent ranking, what it costs per task, what self-hosting a trillion-parameter model actually takes, and what UK and European buyers should check before 27 October.
Table of contents
- What Mistral Announced on 6 October
- Why Mistral Frames Mistral Large 4 as Sovereign AI
- The Cyber Defence Case for Mistral Large 4
- Mistral Large 4 Benchmarks: Mistral’s Own Numbers
- How Mistral Large 4 Ranks on an Independent Index
- What Mistral Large 4 Costs to Use
- What Running Mistral Large 4 Yourself Takes
- Mistral Large 4 Among Open-Weight Rivals
- The Licence Question
- What Mistral Large 4 Means for UK and European Businesses
- What to Watch Before the Weights Ship
- Mistral Large 4: Frequently Asked Questions
- References
What Mistral Announced on 6 October
Mistral’s announcement, “Introducing Mistral Large 4”, opens bluntly: “Unofficially ML4, very officially: le Chonk.” It calls the model “our largest and most capable model to date” and says it “continues to improve rapidly as we refine it”, because reinforcement learning is still running.
The Mistral Large 4 specification sheet
The core facts below come from Mistral’s post unless another source is named. Where outlets disagree, the next section explains why.
| Item | Mistral Large 4 |
|---|---|
| Architecture | Hybrid instruct-and-reasoning mixture-of-experts |
| Total parameters | 1 trillion |
| Active parameters per token | 49 billion |
| Inputs and outputs | Text and images in, text out |
| Languages | More than 160, including every official EU language |
| Training hardware | 3,800 NVIDIA Grace Blackwell GPUs in Europe |
| Context window served | 524,288 tokens (OpenRouter listing) |
| API list price | $1.36 input, $4.18 output, per million tokens |
| Model ID | mistral-large-4-0 |
| Weights | By the end of October 2026 |
Three numbers that differ by source
First, hardware. Mistral’s post says 3,800 Grace Blackwell GPUs, while AFP and Euronews reported “4,000 Nvidia chips” over two months; the larger figure looks like rounding. Second, context. The Decoder reported a one-million-token window, but OpenRouter serves 524,288 tokens and Artificial Analysis measured “about 524,000”. Third, price, which we cover in its own section below.
What Mistral Large 4 replaces
The previous flagship, Mistral Large 3, arrived in December 2025 with 675 billion total and 41 billion active parameters under the Apache 2.0 licence. VentureBeat reports it was trained on 3,000 H200 GPUs. Mistral Large 4 therefore has about 1.48 times the total parameters (1,000 divided by 675) and about 1.2 times the active parameters (49 divided by 41).
Why Mistral Frames Mistral Large 4 as Sovereign AI
“Sovereign” is the word Mistral uses most, and the launch gives it three concrete meanings: where the model was built, where it is served, and who controls it once the weights ship.
Built and served on Mistral’s own compute
“Trained and served on our own compute,” chief executive Arthur Mensch wrote, according to AFP. Mistral’s post says the preview runs on the same European infrastructure used for training. That matters to buyers who worry that a European model served from an American cloud is still subject to American law.
A European region under European law
Mistral says Mistral Large 4 will be available “across multiple regions worldwide, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law”. AFP described it as a “European sovereign region where data stays under EU jurisdiction”. For regulated buyers, that is the headline feature.
Open weights as the final guarantee
The deeper sovereignty argument is the weights. Once released, a bank or a ministry can run Mistral Large 4 on its own servers, with no API dependency at all. Our earlier analysis of private AI for UK businesses explains why that option matters for confidential data.
Who paid for sovereignty
There is an irony worth noting. The Series D that funds Mistral Large 4 was led by Samsung Electronics, a Korean company, as our analysis of the Mistral valuation discussed. Sovereignty here describes where the model runs and who controls the weights, not who owns the cap table.
The Cyber Defence Case for Mistral Large 4
Mistral spends more of its launch on cybersecurity than on any other skill, and the pitch is pointed: closed American models refuse work that defenders need done.
The 82% result
On the Artificial Analysis Cyber Index, Mistral says its model ranks in the top five globally and leads open-weight models built outside China “by a wide margin”. On one test, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral Large 4 scores 82%, “the highest of any model”. It also solves 93% of Cybench, a set of 40 capture-the-flag style exercises.
Refusals as a security risk
Mistral says Claude Opus 5.5 and GPT-6 Astra “score near zero on the same test because they refuse to perform the task”. Its argument is that “defending software often starts with proving that a flaw is real”, and that “losing access to a capability mid-incident can itself become a critical security risk”. Chief scientist Guillaume Lample put it more sharply: the model will help defend against “threat actors that are jailbreaking closed models to perform cyberattacks”.
The safety numbers Mistral published
Mistral also reports that Mistral Large 4 resists 93.3% of attacks on Lakera’s B3 AI Security Benchmark, scores 1.691 out of 2 on the KORA benchmark, and refuses malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm more often than any other open model it tested. Until the weights ship, it is red-teaming with “cybersecurity leaders, vetted partners, and state authorities” using a version with “reduced moderation and expanded cyber capabilities”.
The caveat on dual use
The Decoder points out that Mistral does not explain how the model “reliably tells legitimate vulnerability research apart from attack prep”. Once weights are public, refusals can also be trained out, as the business of removing AI guardrails shows. The cyber strength of Mistral Large 4 cuts both ways, and the open release is what makes it so.
Mistral Large 4 Benchmarks: Mistral's Own Numbers
Mistral published a long list of results. The table below gathers the main ones, with the comparison each figure is set against. All are Mistral’s numbers from its launch post unless marked.
| Benchmark | Mistral Large 4 | Comparison given |
|---|---|---|
| DeepSWE v1.1 | 61.7% | GLM-5.3 61%, DeepSeek V4 Pro 0813 57% (VentureBeat) |
| SWE-Atlas-QnA | 59.4% | Not stated |
| Terminal-Bench 4 | 28.3% | Not stated |
| Coding Agent Index | 49.8% | Ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max |
| AutomationBench (657 workflows) | 59.9% | GLM-5.3 62.2%, Kimi K3 58.3%, DeepSeek V4 Pro 56.7% (The Decoder) |
| Dense 200 visual grounding | 42% | GPT-6 Astra 41% |
| SciCode-Verified | 91.8% | GPT-6 Astra 94.2%, Claude Opus 5 91.3% (The Decoder) |
| Surge AI blind coding review (1 to 5) | 3.74 | Claude Opus 5 4.22, GLM-5.3 3.60, Kimi K3 3.59 |
| AA-Briefcase knowledge work | 1,393 Elo | Ahead of DeepSeek V4 Pro |
Where it leads
Mistral Large 4 is strongest where Mistral focused its training: cyber, visual grounding, and legal and financial documents. Mistral says third-party evaluator Vals.ai found it ahead of GPT-6 Astra on both legal and financial tasks, and that it leads every open model on Harvey’s Legal Agent benchmark.
Where it trails
Coding is competitive rather than leading. GLM-5.3 still beats it on AutomationBench by 2.3 points (62.2 minus 59.9), and in Surge AI’s blind review it sits 0.48 points behind Claude Opus 5 (4.22 minus 3.74). VentureBeat also notes that the live DeepSWE leaderboard shows GLM-5.3 and Kimi K3 near 69% in their best agent set-ups, so a 61.7% preview score “does not establish an outright coding lead”.
How Mistral Large 4 Ranks on an Independent Index
Company benchmarks are chosen by the company. Artificial Analysis ran its own Intelligence Index on the preview on launch day, and its result is the most useful outside view so far.
Bar widths equal the score on the index’s 0 to 100 scale. Scores are as reported by Trending Topics, OfficeChai and The Decoder from the Artificial Analysis chart on launch day.
Best in the West, eighth among open models
The gap to the leader is 19.2 points (57.6 minus 38.4), and the gap to the best open model, Xiaomi’s MiMo-V2.6-Pro, is 7.9 points (46.3 minus 38.4). Trending Topics counted Mistral Large 4 eighth among open-weight models, with seven Chinese models ahead. OfficeChai noted it is the top non-Chinese open model, well clear of Thinking Machines’ Inkling at 25 and Nvidia’s Nemotron 3 Ultra at 23.
A large jump from Large 3
Against its own predecessor the progress is dramatic: about 29 points (38.4 minus roughly 9). It also brings Mistral level with OpenAI’s GPT-6 Luna, which OfficeChai says sits at 38. Mistral’s claim that the preview will keep improving as reinforcement learning continues means this score is a floor, not a final grade.
Long answers, higher bills
There is a cost to that jump. Trending Topics reports that Mistral Large 4 produced about 200 million output tokens across the index, against a median of 81 million for comparable models, roughly 2.5 times more (200 divided by 81). Running the whole index cost about $1,600, or $1.13 per task, up from $0.03 per task for Mistral Large 3.
What Mistral Large 4 Costs to Use
Mistral’s announcement lists $1.36 per million input tokens and $4.18 per million output tokens. OpenRouter, however, lists the preview at exactly half that, $0.68 input and $2.09 output, and The Decoder reported the same preview rate. The safest reading is a launch discount on top of a list price.
Bars are scaled to GPT-6 Astra’s $50, so each width is the price divided by 50. Mistral Large 4’s list price is 8.36% of GPT-6 Astra’s (4.18 divided by 50).
A worked monthly bill
Prices per million tokens are hard to feel, so take a sample workload of 10 million input tokens and 2 million output tokens a month, a modest internal assistant. The table multiplies each listed price out. It ignores caching and batch discounts, which change every row.
| Model | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Mistral Large 4, list | 10 x $1.36 = $13.60 | 2 x $4.18 = $8.36 | $21.96 |
| Mistral Large 4, preview | 10 x $0.68 = $6.80 | 2 x $2.09 = $4.18 | $10.98 |
| DeepSeek V4 Pro 0813 | 10 x $0.66 = $6.60 | 2 x $1.98 = $3.96 | $10.56 |
| Kimi K3 | 10 x $0.99 = $9.90 | 2 x $14.00 = $28.00 | $37.90 |
| Claude Opus 5.5 | 10 x $4.00 = $40.00 | 2 x $20.00 = $40.00 | $80.00 |
| GPT-6 Astra | 10 x $10.00 = $100.00 | 2 x $50.00 = $100.00 | $200.00 |
Price per answer, not per token
Per token, Mistral Large 4 is far cheaper than the closed leaders. But a model that writes 2.5 times as many tokens to answer the same question narrows that gap. Test it on your own prompts and compare the cost per finished task, which is how Artificial Analysis reached its $1.13 figure. Our LLM API pricing guide walks through that method.
What Running Mistral Large 4 Yourself Takes
The weights are the point of an open model, but a trillion parameters is a lot of memory. Mistral has not yet said what precision the weights will ship in, so the figures below cover the common formats.
The memory arithmetic
Each parameter takes 2 bytes at 16-bit precision, 1 byte at 8-bit and half a byte at 4-bit. For 1 trillion parameters that gives about 2 TB, 1 TB and 500 GB of weights. A server with eight 80 GB H100 GPUs holds 640 GB (8 x 80), and one with eight 141 GB H200 GPUs holds 1,128 GB (8 x 141).
| Precision | Weights size | 8 x H100 (640 GB) | 8 x H200 (1,128 GB) |
|---|---|---|---|
| 16-bit | About 2,000 GB | Does not fit | Does not fit |
| 8-bit | About 1,000 GB | Does not fit | Fits, about 128 GB spare |
| 4-bit | About 500 GB | Fits, about 140 GB spare | Fits, about 628 GB spare |
The spare memory has to hold the conversation cache, which grows with context length, so a long-context deployment needs more headroom than these subtractions suggest.
Active parameters help speed, not memory
Because only 49 billion parameters work on each token, Mistral Large 4 does roughly the per-token work of a 49-billion-parameter dense model, which helps explain the 116 tokens a second Artificial Analysis measured, according to Trending Topics. But every expert must still sit in memory. Self-hosting is therefore a multi-GPU server project, not a workstation one.
Who will actually self-host
Realistically, banks, defence suppliers, telecoms operators and government agencies will run the full model. Smaller firms are more likely to use Mistral’s European region, or wait for the “new generation of specialized and optimized Mistral models” that Mistral says Mistral Large 4 will be the foundation for.
Mistral Large 4 Among Open-Weight Rivals
Mistral Large 4 lands in a busy fortnight for open models. Two other Western releases arrived days apart, and their basic specs make a useful comparison.
| Model | Total / active parameters | Context | Licence | Weights |
|---|---|---|---|---|
| Mistral Large 4 | 1T / 49B | 524,288 served | Not yet published | End of October |
| Reflection Beam | 501B / 23B | 1 million | Apache 2.0 | Later in October |
| Aleph Alpha Kolibri | 78.1B / 3.46B | 1,048,576 | Apache 2.0 | Released 3 October |
| Mistral Large 3 | 675B / 41B | 262,144 | Apache 2.0 | Released December 2025 |
Against Reflection Beam
America’s answer arrived the day before. Reflection Beam is half the size, text-only, and promises Apache 2.0. On DeepSWE, VentureBeat’s comparison put Beam at 44% against 62% for Mistral Large 4, but Beam’s real test, like Mistral’s, comes when the weights are public.
Against Aleph Alpha Kolibri
Germany’s Aleph Alpha Kolibri is a different tool: small, fast, English and German only, and already downloadable. Kolibri runs on one server; Mistral Large 4 needs a rack-sized budget. Many European buyers will end up using both for different jobs.
The Licence Question
The licence may matter more than any benchmark. Mistral Large 3 shipped under Apache 2.0, which allows commercial use with almost no conditions. For Mistral Large 4, VentureBeat reports a custom Mistral licence rather than Apache 2.0, while The Decoder says the licence will be published with the weights.
Why it matters
A custom licence can restrict commercial use above a revenue threshold, ban certain uses, or require attribution. Until the text appears, nobody should plan a product on Mistral Large 4 weights. If it turns out restrictive, the preview’s open-weight framing will deserve a footnote.
The AI Act angle
Open weights do not exempt the largest models from the EU AI Act. General-purpose models presumed to carry systemic risk, which includes those trained with more than 10^25 floating-point operations, keep extra duties on evaluation and incident reporting. Mistral has not published training compute for Mistral Large 4, so where it sits against that line is not yet public.
What Mistral Large 4 Means for UK and European Businesses
For most organisations, the launch is less about benchmarks and more about options. Three practical questions follow.
Decide where the data may go
If your data must stay under EU or UK control, Mistral’s European region and the coming weights give you two routes that US-hosted models do not. Write down which data classes may go to an external API, which must stay on your own servers, and which can use either.
Pilot on your own tasks
Run Mistral Large 4 against the model you use today on 50 to 100 real tasks from your own work. Measure accuracy, cost per finished task and response length. The verbosity Artificial Analysis found will show up in your bill, so record tokens as well as answers.
Treat the cyber skills as a policy question
If your security team wants Mistral Large 4 for vulnerability work, set rules first: who may use it, on which systems, and how outputs are logged. Our AI strategy work builds exactly that kind of usage policy before a pilot starts.
What to Watch Before the Weights Ship
The preview is a three-week window before the weights arrive. These are the signals that will decide whether Mistral Large 4 lives up to its launch.
The licence text
The first thing to read on release day. Commercial terms, use restrictions and any revenue thresholds decide whether Mistral Large 4 is really open for business use.
Independent re-tests
Artificial Analysis will rerun its index as reinforcement learning continues, and the DeepSWE leaderboard should list the final model. If the score climbs from 38.4 towards the Chinese leaders, Mistral’s “no sign of saturation” claim will look justified.
The weight format and size
The precision the weights ship in sets the hardware bill. An official 8-bit or 4-bit release would bring Mistral Large 4 within reach of a single eight-GPU server.
Mistral Large 4: Frequently Asked Questions
What is Mistral Large 4?
Mistral Large 4 is Mistral AI’s largest model, a 1-trillion-parameter mixture-of-experts model with 49 billion active parameters. It takes text and images as input and answers in text, in more than 160 languages.
When will the Mistral Large 4 weights be released?
Mistral says “by the end of the month”. AFP and VentureBeat reported 27 October 2026. The preview API is available now through Mistral’s platform.
How much does Mistral Large 4 cost?
Mistral lists $1.36 per million input tokens and $4.18 per million output tokens. OpenRouter currently lists the preview at half that, $0.68 and $2.09.
Is Mistral Large 4 open source?
It is open-weight: the weights will be downloadable. The licence has not been published yet, and VentureBeat reports it will be a custom Mistral licence rather than Apache 2.0.
Why does Mistral call it sovereign AI?
Because it was trained and is served on Mistral’s own European infrastructure, can run in a European region operated under European law, and can be self-hosted once the weights ship.
References
Introducing Mistral Large 4 (Mistral AI)
France’s Mistral unveils latest model in sovereign AI push (AFP via France 24)
France’s Mistral unveils latest model in sovereign AI push (Tech Xplore)
What to know about Mistral’s new model in the US-China open AI race (Euronews)
Mistral debuts Large 4 Le Chonk (VentureBeat)
Mistral Large 4 trails China’s open-weight leaders on Artificial Analysis (Trending Topics)
Mistral Large 4 becomes top non-Chinese open model on Artificial Analysis (OfficeChai)
Mistral launches Large 4 preview with 1T parameters (TestingCatalog)
Mistral Large 4 on OpenRouter (OpenRouter)
Artificial Analysis Intelligence Index (Artificial Analysis)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
