Mistral AI pulled back the curtain on Mistral Large 4 on October 6, 2026, and by October 15 the model, internally nicknamed “Le Chonk,” had become the clearest signal yet that Europe intends to field its own frontier-scale AI model rather than simply license one from Silicon Valley or Beijing. According to Mistral’s own announcement, Large 4 is a mixture-of-experts system built around roughly 1.05 trillion total parameters, with a public API preview now live through Mistral Studio while the promised open-weight release is still pending.
The timing matters. Mistral Large 4 landed in the same stretch of October that saw GPT-6.1 Sol and Google’s Gemini 4 Argon trade blows over million-token context windows, and just days after Nvidia pushed out its own Nemotron 3.5 Super VL model. Mistral Large 4 enters a market where every major lab is racing to claim the biggest context window, the cheapest token price, or the most defensible sovereignty story. Mistral is leaning hard into the third option.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Mistral Large 4: The Headline Numbers
Mistral Large 4 is a granular mixture-of-experts model, meaning only a slice of its total parameters activates for any given token rather than the full network firing at once. Mistral’s announcement puts the total parameter count at approximately 1.05 trillion, with 52 billion active parameters per forward pass, though several secondary outlets tracking the release, including Unite.AI, cite a slightly lower figure of 49 billion active parameters. Some reports also point to an additional 1.6-billion-parameter vision encoder bolted onto the core model, though that detail has not yet appeared in a full technical report from Mistral itself.
Whichever active-parameter figure proves accurate, the gap between 1.05 trillion total and roughly 50 billion active is the entire point of a mixture-of-experts design: it lets Mistral claim a trillion-parameter-class model while keeping inference costs closer to what a mid-size dense model would require. That tradeoff is also why Mistral is pitching Large 4 as hybrid instruct-and-reasoning model rather than a pure chat model, a framing meant to put it in the same conversation as reasoning-heavy releases from OpenAI and DeepSeek.
A Million-Token Context Window, With an Asterisk
Mistral’s announcement lists a 1-million-token context window for Large 4, matching the context ceiling that Google and OpenAI have both built into their newest releases. In practice, the preview API currently appears to cap out at 512,000 tokens, roughly half the advertised maximum, according to coverage of the rollout. That gap between announced ceiling and shipped limit is common in early previews, but it means developers testing Large 4 against its stated spec today will not yet see the full window in production.
A 1-million-token window, once fully available, would let Mistral Large 4 ingest entire codebases, long legal contracts, or hours of transcribed audio in a single pass without chunking. That is table stakes among 2026’s frontier models, but it is a meaningful jump for Mistral specifically, whose earlier Large-series models topped out at far smaller windows.
Mistral Large 4 Pricing: What the API Actually Costs
Mistral listed API pricing for Large 4 at $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens, according to pricing details compiled by Eden AI. Mistral reportedly sweetened the launch with a 50 percent discount for the first two weeks, which would put effective pricing closer to $0.68 per million input tokens and $2.09 per million output tokens during that window. A blended-cost estimate circulating among benchmark trackers puts an average cost near $2.06 per million tokens, though that figure depends heavily on the assumed mix of input versus output traffic and should not be read as official list pricing.
That pricing places Mistral Large 4 above the cheapest frontier-adjacent options on the market right now. Google’s Gemini 4 Argon opened at an introductory $2 per million input tokens, and multiple budget-oriented releases this quarter, including Mistral’s own smaller models, have pushed per-token costs lower still. Large 4 is not positioned as the cheap option. It is positioned as the European option.
Mistral Large 4 Technical Specifications at a Glance
| Spec | Mistral Large 4 |
|---|---|
| Announcement date | October 6, 2026 |
| Architecture | Granular mixture-of-experts, hybrid instruct/reasoning |
| Total parameters | Approximately 1.05 trillion |
| Active parameters | 49-52 billion (figures vary by source) |
| Context window | 1 million tokens announced; 512,000 tokens exposed in preview |
| Modalities | Text and image input; text output |
| Input price | $1.36 per million tokens |
| Cached input price | $0.14 per million tokens |
| Output price | $4.18 per million tokens |
| Training infrastructure | 3,800 Nvidia Grace Blackwell GPUs, Mistral’s own European datacenters |
| License | Not yet finalized; open-weight release promised |
| Availability | Public preview via Mistral Studio and the Mistral API |
Where Mistral Large 4 Lands on Early Benchmarks
Mistral has not published a single controlled benchmark table pitting Large 4 against GPT-6, Gemini 4 Argon, Claude, and DeepSeek under identical conditions, so any comparison right now is necessarily a patchwork of independent test runs rather than an apples-to-apples leaderboard. On the Artificial Analysis Intelligence Index, Mistral Large 4 scores approximately 38, putting it roughly level with GPT-6 Luna, also at 38, and just behind DeepSeek V4.1 Flash at 39, based on an independent comparison cited by benchmark trackers. On more task-specific tests, Mistral reported a 61.7 percent score on DeepSWE v1.1, 28.3 percent on Terminal-Bench 4, and 59.9 percent on AutomationBench.
The BenchLeader aggregator, which pools results across a wide range of public evaluations, places Mistral Large 4 at a composite score of 73, with a margin of roughly plus or minus 5.1, ranking it 131st out of 767 models tracked on that site. That is a respectable mid-pack position among every model BenchLeader has ever logged, though it is far from a frontier-topping score.
Mistral Large 4 vs GPT-6 Luna vs DeepSeek V4.1 Flash
| Model | Artificial Analysis Intelligence Index | Notes |
|---|---|---|
| DeepSeek V4.1 Flash | 39 | Slightly ahead of Mistral Large 4 in this comparison |
| GPT-6 Luna | 38 | Statistically even with Mistral Large 4 |
| Mistral Large 4 | ~38 | Up sharply from its predecessor |
| Mistral Large 3 (prior gen) | 9.3 | Previous Mistral flagship, for reference |
That last row is the number that matters most for Mistral internally. A jump from an Artificial Analysis score of roughly 9.3 for Large 3 to approximately 38 for Large 4 is one of the larger single-generation improvements any lab has logged on that index this year, even if the absolute score still trails the leading US and Chinese systems. Tom’s Hardware framed the result bluntly: independent testing ranks Mistral Large 4 as the strongest AI model built outside the United States and China, while noting that some Chinese open-weight systems still outperform it on the same tests.
Multimodal, But Not Fully
Mistral Large 4 accepts text and image input, including document understanding, and returns text output. That is a real step up from a text-only model, but it falls short of the full text-image-audio-video packages that some competitors have started shipping. There is no confirmed audio input or output, and no confirmed video support, in the material Mistral has published so far. If Mistral intends to catch up on those modalities, it has not said so publicly as of this release.
The Open-Weight Promise Nobody Can Download Yet
Mistral has built its reputation partly on open-weight releases, and the company is describing Large 4 as open-weight in its launch messaging. The catch is that, as of October 15, the weights are not yet downloadable and no final license has been published. Reports point to a full open-weight release by the end of October 2026, with one report naming October 27 specifically, though that date has not been confirmed in Mistral’s primary announcement. Until Mistral publishes both the weights and the license terms, Large 4 should be treated as an API-only preview rather than a genuinely open model, regardless of how it is being marketed.
That distinction matters for anyone weighing Large 4 against other open-weight contenders in the ai-models space right now, including the comparison field covered in GLM 5.3 Prime, Qwen3.8 Max, and Step 5, all of which already have usable public weights. Mistral is asking developers to commit to its API today on the promise of open weights tomorrow.
Why Mistral Trained in Its Own European Datacenters
Mistral’s announcement states that Large 4 was “trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe.” That line is doing a lot of strategic work. Every major AI lab training a trillion-parameter-class model from outside Europe has, by necessity, relied on US cloud infrastructure, Chinese domestic chips, or some mix of both. Mistral is positioning Large 4 as proof that a European company can train a frontier-adjacent model on European soil, using European-operated infrastructure, without routing training runs through Microsoft Azure, Amazon Web Services, or Google Cloud.
That sovereignty pitch lines up with a broader European push toward AI independence that has also shown up in Nvidia’s own European initiatives, including its backing of the RAISE research effort aimed at expanding European AI research capacity. Whether training location alone satisfies EU AI Act obligations around documentation, risk classification, and transparency is a separate question that Mistral’s announcement does not fully answer. The company has not published, as of this writing, a definitive statement on how Large 4 will be classified under the EU AI Act’s rules for general-purpose models, nor what copyright and training-data documentation it will provide alongside the eventual open-weight release.
From Large 2 and Large 3 to Large 4: A Steep Climb
Mistral’s prior flagship, Large 3, scored around 9.3 on the Artificial Analysis Intelligence Index, a number that looks almost quaint next to Large 4’s roughly 38. The available public material does not give a full architecture-by-architecture comparison between the two generations, but the scale of the jump, combined with the move to a trillion-parameter mixture-of-experts design and a stated million-token context window, signals that Mistral treated Large 4 as a from-scratch rebuild rather than an incremental update. That is consistent with the company’s own description of training the model from scratch rather than fine-tuning a prior checkpoint.
For a company that has spent the better part of three years positioning itself as Europe’s answer to OpenAI and Anthropic, a flagship model that roughly quadruples its predecessor’s benchmark score is a meaningful proof point, even before the open-weight release lands.
Competitive Landscape: A Crowded October for Frontier Models
Mistral Large 4 did not launch into a quiet market. The same week saw Google roll out Gemini 4 Argon at $2 per million input tokens, with coverage of its arrival inside Android Studio’s coding workflows underlining how fast the major labs are shipping. Nvidia separately released its Nemotron 3.5 family, including the 120-billion-parameter Super VL variant, aimed squarely at developers who want a smaller, faster multimodal model rather than a trillion-parameter giant. Anthropic has kept pace on pricing with its Claude Haiku 5.5 launch, which undercut its own prior generation by a wide margin.
Against that backdrop, Mistral Large 4’s benchmark position, roughly level with GPT-6 Luna and just behind DeepSeek V4.1 Flash on the Artificial Analysis index, reads less as a knockout release and more as Mistral buying itself a seat at the table. The company is not claiming to have the best model in the world. It is claiming to have the best model trained and operated entirely in Europe, which is a different, narrower, and arguably more defensible claim.
Market Impact: What’s Actually Riding on This for Mistral
Mistral AI is privately held, so there is no stock price to move on this news, and no confirmed new funding round, valuation change, or investor statement has been tied directly to the Large 4 launch. What Large 4 does affect is Mistral’s standing as the default European alternative for enterprises and governments that want contractual or regulatory assurance that their AI workloads are not running on US or Chinese infrastructure. That is a smaller market than the general-purpose consumer chatbot business OpenAI and Google are fighting over, but it is a market where Mistral has comparatively little direct competition, and Large 4 is the company’s strongest card yet to play in it.
The pricing structure, at $1.36 per million input tokens and $4.18 per million output tokens before any discount, also signals that Mistral is not trying to win on cost. It is trying to win on provenance. That is a bet that European data-residency and sovereignty requirements will keep growing as a procurement criterion, a trend that shows up elsewhere in the European AI infrastructure build-out that outlets have tracked throughout 2026.
Historical Context: Europe’s Long Road to a Frontier Model
Mistral AI was founded in 2023 by former DeepMind and Meta researchers specifically to give Europe a homegrown alternative in a field dominated by US labs and, increasingly, Chinese open-weight releases. Its early models, including the original Mistral 7B, won attention for competitive performance at small scale, well before the company had the compute budget to chase trillion-parameter territory. The jump to Large 4, trained on thousands of Nvidia’s newest Grace Blackwell chips inside Mistral’s own datacenters, represents the clearest evidence yet that the company has closed at least part of the infrastructure gap that separated it from Silicon Valley’s biggest labs. It also arrives amid a broader wave of open-weight competition, a trend tracked in coverage of how open-source models are reshaping the AI frontier more generally.
What Happens Next: Predictions for Mistral Large 4
- The open-weight release, if it lands around the reported late-October target, will be the real test of whether Mistral Large 4 earns developer adoption beyond the API preview, since many enterprises and researchers specifically want downloadable weights for on-premises or air-gapped deployment.
- Expect Mistral to publish a fuller technical report once weights ship, likely clarifying the exact active-parameter count and resolving the 49-billion-versus-52-billion discrepancy currently circulating in secondary coverage.
- The 512,000-token preview cap will likely expand toward the advertised 1-million-token ceiling as Mistral scales out serving infrastructure, mirroring how other labs have phased in context-window increases after initial launch.
- Pricing will almost certainly shift once the 50 percent launch discount expires, and Mistral will face pressure to cut list prices further if Gemini 4 Argon’s $2-per-million-token entry point holds up against competitive responses from OpenAI and DeepSeek.
- European governments and regulated enterprises are the most likely early adopters, given Mistral’s sovereignty pitch, ahead of broader consumer-facing adoption that would require Mistral to close the remaining benchmark gap with GPT-6 and DeepSeek’s top models.
Frequently Asked Questions
When was Mistral Large 4 released?
Mistral announced Large 4 on October 6, 2026, opening it as a public-preview API release through Mistral Studio.
How many parameters does Mistral Large 4 have?
Mistral Large 4 has approximately 1.05 trillion total parameters in a mixture-of-experts architecture, with reported active-parameter figures of either 49 billion or 52 billion depending on the source.
Is Mistral Large 4 open-weight?
Mistral describes Large 4 as open-weight, but as of October 15, 2026, the weights were not yet downloadable and no final license had been published. A full open-weight release is reportedly expected by the end of October 2026.
How much does Mistral Large 4 cost to use via the API?
Listed pricing is $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens, with a reported 50 percent launch discount for the first two weeks.
How does Mistral Large 4 compare to GPT-6 and DeepSeek?
On the Artificial Analysis Intelligence Index, Mistral Large 4 scores approximately 38, roughly matching GPT-6 Luna at 38 and trailing DeepSeek V4.1 Flash at 39. Independent testing cited by Tom’s Hardware ranks it as the strongest model built outside the US and China, while noting some Chinese open-weight models still score higher.
What is Mistral Large 4’s context window?
Mistral’s announcement lists a 1-million-token context window, though the current API preview reportedly exposes only up to 512,000 tokens.
Does Mistral Large 4 support images, audio, or video?
Mistral Large 4 supports text and image input, including document understanding, with text-only output. There is no confirmed audio or video support in the material Mistral has published so far.
Why is Mistral emphasizing that Large 4 was trained in Europe?
Mistral says Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs inside the company’s own European datacenters, a detail central to its pitch that Europe can field a frontier-adjacent AI model without relying on US or Chinese infrastructure.
