Mistral AI pushed a trillion-parameter model into developers’ hands on October 6, 2026, and gave it a name that sounds more like a nickname for a large dog than a frontier AI system. Mistral Large 4, internally abbreviated ML4 and branded in the announcement as “Le Chonk,” is a 1.05 trillion-parameter open-weight multimodal mixture-of-experts model that activates roughly 52 billion parameters per token, according to the company’s own October 2026 launch post and reporting from MarkTechPost. The Paris-based lab is positioning it as the strongest open-weight model built outside China, and it is leading with cybersecurity benchmarks rather than the usual chatbot leaderboard flex.
The timing matters. Mistral Large 4 lands the same week that European policymakers are leaning harder on “AI sovereignty” language, and just months after Reflection AI’s $25 billion bet to take on China’s open models made headlines for a similar pitch from a different continent. This piece walks through what Mistral actually shipped, how it compares to the open-weight competition, what it costs today, and what is still missing until the weights themselves are released later this month.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Mistral Large 4 Actually Is: Architecture and Specs
Mistral Large 4 is a mixture-of-experts (MoE) model, meaning it does not activate its full parameter count on every request. According to Mistral’s own announcement and reporting from TestingCatalog, the model totals roughly 1.05 trillion parameters but only activates about 52 billion of them per token, a figure Mistral AI itself confirmed in its October 2026 model documentation. That routing design is what keeps inference costs manageable despite the headline parameter count – the same basic approach OpenAI, DeepSeek, and Alibaba have all adopted for their own largest releases.
Mistral’s own post describes the model in blunt terms: “ML4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters,” the company said in its October 6 launch announcement. Reporting also attributes to the company a description of the model as an “open-weight hybrid instruct-and-reasoning MoE with multimodal input,” built to unify instruction-following, reasoning, and agentic tool use inside one system rather than splitting them across separate models.
Mixture-of-Experts, in Plain Terms
Think of an MoE model as a large team of specialists with a router that decides which specialists handle each request. Instead of every parameter firing for every token, a gating network picks a small subset of “expert” sub-networks to process each piece of input. That is how Mistral Large 4 can carry a trillion-plus parameters in storage while behaving, computationally, closer to a 52-billion-parameter dense model at inference time, per Mistral AI’s own October 2026 figures. It is the same trick that let Mixtral 8x7B – one of Mistral’s earlier releases – punch above its weight back in 2023 and 2024.
Multimodal Input and the 1-Million-Token Context Window
Beyond text, Mistral Large 4 ships with a 1.6 billion-parameter vision encoder bolted onto the main network, giving it native image understanding rather than a bolted-on captioning layer, per reporting from AI TL;DR and Mistral AI’s own October 2026 model documentation. That same model card confirms the 1-million-token context window, putting it in the same long-context tier as the largest releases from Google and Anthropic this year. For enterprise users, that combination is the actual pitch: a model that can read a codebase, a contract, and a screenshot in the same request without switching tools.
Inside the Training Run: Thousands of Blackwell GPUs, Two Months, 160+ Languages
Mistral says it trained Large 4 from scratch in roughly two months, running the job on its own European data center capacity rather than renting external cloud compute. Reports on the exact cluster size differ slightly: TechCrunch reported on October 6, 2026 that training ran on roughly 4,000 Nvidia GPUs, while OfficeChai’s coverage and others cite a slightly lower count near 3,800. Either way, it is a sovereign-compute story as much as a model story – Mistral has spent much of 2026 building out its own data centers in France and Sweden specifically so it is not dependent on US hyperscaler capacity for training runs like this one.
The training data spanned more than 160 languages, including every official language of the European Union, according to reporting on the launch from Capital Brief. That language coverage is a deliberate differentiator: most frontier labs optimize overwhelmingly for English and Mandarin benchmarks, and Mistral has built its commercial identity around serving European governments and enterprises that need strong performance in French, German, Polish, and smaller EU languages that larger US and Chinese labs tend to treat as an afterthought.
The Cybersecurity Benchmark Numbers Everyone Is Citing
Mistral did not lead its announcement with a generic reasoning or coding score. It led with cybersecurity. The company reports that Mistral Large 4 scored 93% on Cybench, a benchmark built from roughly 40 exercises drawn from real security competitions, and 82% on CyberGym-E2E, a reproduce-and-patch style test that checks whether a model can both find and fix a vulnerability rather than just describe it.
What makes that framing notable is the comparison point Mistral chose: the company says several closed frontier models score near zero on these same cybersecurity tasks, not because they lack the technical capability, but because their safety filters refuse to engage with offensive-security-flavored prompts at all. Mistral is effectively arguing that an open-weight model tuned for defensive security work – vulnerability research, patch generation, red-team simulation – fills a gap that safety-conservative closed models leave open. That is a deliberate pitch to “cybersecurity leaders,” a group reporting says received early preview access alongside government and state authorities before the wider developer rollout.
It is also a pitch with obvious dual-use tension. A model that scores 93% on a benchmark built from offensive security exercises is, by definition, a model that is good at finding and exploiting vulnerabilities – the same skill set that defenders and attackers both need. Mistral’s bet is that restricting early access to vetted security professionals and government bodies manages that risk while still proving the model’s value before the weights ship publicly.
Mistral Large 4 Spec Sheet
| Spec | Mistral Large 4 (“Le Chonk”) |
| Total parameters | 1.05 trillion |
| Active parameters per token | 49 billion |
| Architecture | Mixture-of-experts, hybrid instruct + reasoning |
| Vision encoder | 1.6 billion parameters (native multimodal) |
| Context window | 1 million tokens |
| Training hardware | ~3,800-4,000 Nvidia Grace Blackwell GPUs |
| Training duration | ~2 months, from scratch |
| Languages trained on | 160+, including all official EU languages |
| Preview API pricing | $1.36 / 1M input tokens, $4.18 / 1M output tokens |
| Open weights release | Planned late October 2026 (reports vary between Oct. 27 and Oct. 31) |
How Mistral Large 4 Stacks Up Against DeepSeek, Qwen, and Kimi
Mistral is not shy about naming its competition. The company’s own benchmarking pits Large 4 directly against DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, and GLM-5.3 – all open-weight models released by Chinese labs over the past year. That comparison set tells its own story: the open-weight frontier in 2026 is overwhelmingly a China-versus-everyone-else race, and Mistral is positioning itself as the one credible non-Chinese entrant still competing at the top of that table, a framing echoed by CNBC’s coverage of CEO Arthur Mensch’s comments on the launch.
Where Mistral Claims the Lead
Mistral says Large 4 is the strongest open-weight model developed outside China “by a substantial margin,” and reporting indicates it scores best among open models on SciCode-Verified, a benchmark for scientific coding tasks, while also outperforming DeepSeek V4 Pro and Qwen3.8 Max on a combined coding-agent index that blends several agentic-coding tests. Mistral AI also reports that Large 4 scored 59.9% on AutomationBench’s 657 workflows, and on visual grounding the company’s own page says “we find ML4 to be one of the most capable models we tested, for instance surpassing GPT-6-Astra on Dense 200 (42% vs 41%).”
Where the Gap Narrows
The margin against its closest Chinese rivals is thinner than Mistral’s marketing language suggests. Analysis from AI TL;DR notes that the numerical distance between Large 4 and Qwen3.8 Max, or between Large 4 and DeepSeek V4 Pro, is small enough that a sweeping “best overall” ranking claim deserves skepticism – the displayed uncertainty bands and workload-specific results matter more than a few tenths of a percentage point on any single leaderboard. Qwen3.8 Max, for context, is a larger model on paper at 2.4 trillion total parameters with 95 billion active, while DeepSeek V4 Pro runs 1.65 trillion total parameters with a matching 49 billion active parameters and sparse attention, having launched in August 2026.
Open-Weight Frontier Models Compared
| Model | Total Params | Active Params | Release |
| Mistral Large 4 (“Le Chonk”) | 1.05T | 49B | Oct. 2026 (preview); weights late Oct. 2026 |
| DeepSeek V4 Pro | 1.65T | 49B | Aug. 2026 |
| Qwen3.8 Max | 2.4T | 95B | 2026 |
| Kimi K2 | 1T | 32B | Jul. 2025 |
For readers tracking the open-source model race more broadly, this table is worth reading alongside our recent comparison of MiniMax M3.1, GPT-6 Luna, and DeepSeek V4.1 Flash, which covers a different slice of the same pricing and performance tradeoffs playing out across the industry this quarter.
Pricing and Availability: What Developers Get Today
Mistral Large 4 went live in preview on October 6, 2026, with API access priced at $1.36 per million input tokens and $4.18 per million output tokens. That preview pricing sits well above what Mistral charges for its smaller models, reflecting both the compute cost of serving a trillion-parameter MoE and the limited-access nature of a preview launch. Some reports note the company has floated a lower Standard-tier rate for later in the rollout, though preview pricing is the only figure currently confirmed by the launch itself.
Crucially, “open-weight” does not yet mean “downloadable.” The weights themselves – the actual model files developers would need to self-host Large 4 – are not shipping on day one. Mistral says it plans to release them by the end of October 2026, after what it describes as further safety testing; Le Monde reported that the weights would be released by October 27, 2026, with other outlets citing dates ranging up to October 31. Until then, the only way to use Mistral Large 4 is through Mistral’s own hosted API, which somewhat undercuts the open-weight framing in the short term, even if the end state is a genuinely downloadable model.
That staged rollout mirrors a pattern readers will recognize from other recent model launches – compare it to how Z-Image Turbo’s local setup required waiting for a public weights drop before self-hosting was realistic, or how GLM-5.3-Flash split its rollout between an immediate API and a delayed self-host option.
Calling the Preview API
Mistral’s hosted API follows the same request pattern as its existing Large and Medium model lines, so developers already using the Mistral SDK should find the preview familiar. An illustrative request, based on Mistral’s established API conventions, looks like this:
curl https://api.mistral.ai/v1/chat/completions \
-H "Authorization: Bearer $MISTRAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistral-large-4",
"messages": [
{"role": "user", "content": "Summarize this incident report and flag any CVEs mentioned."}
]
}'Note that the exact model identifier string may change between preview and general availability, so developers should confirm the current name against Mistral’s own API documentation before shipping production code against it.
Why Mistral Is Betting on Cybersecurity, Not Just Chat
Every major lab is chasing coding and agentic benchmarks right now, which makes Mistral’s cybersecurity-first framing a genuine product decision rather than a marketing afterthought. Security work is a market where European firms face a specific regulatory and sourcing pressure: many cannot legally or contractually route sensitive vulnerability research through US-hosted or Chinese-hosted models. A European lab offering a model that scores competitively on Cybench and CyberGym-E2E, hosted on European infrastructure, is a direct answer to that sourcing problem – not a theoretical one.
That also explains the staged access list: developers, cybersecurity leaders, and state authorities first, broader public access later. A model this capable at offensive-security tasks invites exactly the kind of misuse scrutiny that slower, more controlled rollouts are designed to manage, and Mistral appears to be front-loading that caution rather than treating it as a post-launch cleanup problem.
Mistral’s Road to Le Chonk: From a Paris Startup to a €21 Billion Valuation
Mistral AI was founded in April 2023 in Paris by Arthur Mensch, Timothée Lacroix, and Guillaume Lample, all of whom came out of DeepMind and Meta’s AI research groups. The company built its early reputation on open releases – Mistral 7B and the Mixtral 8x7B mixture-of-experts model were both widely adopted by developers specifically because they were downloadable and modifiable, in contrast to the closed-API-only approach of OpenAI and Anthropic at the time.
That open-first identity has persisted through a rapid string of funding rounds. Mistral closed a Series D in September 2026 led by Samsung Electronics, raising roughly €3 billion and pushing its valuation above €21 billion – the highest of any European AI company. The model line has grown alongside the funding: Mistral Large 3, Mistral Medium 3.5, and Mistral Small 4 all shipped earlier in 2026, with Large 4 now sitting at the top of that lineup as the company’s flagship frontier release.
Market Impact and Investor Reaction
For enterprise buyers, Mistral Large 4 adds a third credible open-weight option alongside the Chinese model families and whatever Meta and smaller labs ship next, at a moment when procurement teams are increasingly asked to document where their AI inference actually runs. European customers get a model trained and served on European infrastructure, a story Mistral has been telling investors for two years and can now point to as reflected in product shipping.
The broader open-weight market is also clearly commoditizing at the top end. With DeepSeek V4 Pro, Qwen3.8 Max, Kimi K2, GLM-5.3, and now Mistral Large 4 all trading benchmark leads within a few percentage points of each other, enterprise buyers increasingly have to weigh price, hosting location, context window, and language support rather than raw leaderboard position. That shift favors labs like Mistral that can differentiate on something other than a marginal benchmark win – in this case, European data residency and a cybersecurity specialization.
What Mistral Is Saying About Le Chonk
Mistral’s own language from the October 6 launch post frames the release plainly: “Today, we’re launching a public preview of Mistral Large 4,” the company said, adding the informal branding directly in the same announcement: “Unofficially ML4, very officially: le Chonk.” On the model’s positioning, Mistral’s stated ambition is unambiguous – the company says “ML4 pushes the frontier of open-weight performance,” describing the system as an “open-weight hybrid instruct-and-reasoning MoE with multimodal input” built to unify instruction-following, reasoning, and agentic capability in one release.
5 Predictions for What Happens Next
- The late-October weights release becomes the real test. Preview benchmarks are company-reported; independent evaluation only becomes possible once the weights are downloadable, likely between October 27 and 31, 2026.
- Rivals will respond with their own cybersecurity-tuned variants. Expect DeepSeek, Qwen, and Kimi’s maintainers to publish competing Cybench and CyberGym scores within weeks of Mistral’s numbers circulating.
- Enterprise European buyers move first. Government and regulated-industry customers who need EU-hosted inference are the most likely early adopters, ahead of general developer uptake.
- Pricing will compress. List pricing of $1.36/$4.18 per million tokens is unlikely to hold once the model exits preview and competitive pressure from Chinese open-weight pricing sets in.
- Scrutiny of the cybersecurity framing will intensify. A model this capable at offensive security benchmarks will draw attention from policy researchers questioning how access controls hold up once weights are public.
Frequently Asked Questions
What is Mistral Large 4?
It is Mistral AI’s newest flagship model, a 1.05 trillion-parameter open-weight multimodal mixture-of-experts system announced October 6, 2026, and nicknamed “Le Chonk.”
Why is it called “Le Chonk”?
Mistral used the nickname directly in its own launch announcement, alongside the formal abbreviation ML4, as an informal reference to the model’s trillion-parameter size.
Can I download Mistral Large 4’s weights today?
Not yet. As of October 6, 2026, only hosted API access is available. Mistral says open weights will follow by the end of October 2026, with reports citing dates between October 27 and 31.
How much does the Mistral Large 4 preview API cost?
The list price confirmed at launch is $1.36 per million input tokens and $4.18 per million output tokens.
How does Mistral Large 4 compare to DeepSeek V4 Pro and Qwen3.8 Max?
It is smaller in total parameters than both (1.05T versus 1.65T and 2.4T respectively) but matches DeepSeek V4 Pro’s 49B active-parameter count. Reported benchmark margins between the three are narrow enough that no single model holds a decisive lead across the board.
What makes Mistral Large 4’s cybersecurity benchmarks notable?
Mistral reports 93% on Cybench and 82% on CyberGym-E2E, scores the company says beat several closed frontier models that refuse offensive-security-style prompts outright rather than attempting them.
Who can access Mistral Large 4 right now?
The preview rolled out first to developers, cybersecurity professionals, and state/government authorities, ahead of a broader public release later in October.
What context window does Mistral Large 4 support?
1 million tokens, alongside native multimodal input through a 1.6 billion-parameter vision encoder.
