On October 6, 2026, Mistral AI opened a public preview of Mistral Large 4, a multimodal mixture-of-experts model the company nicknames “Le Chonk.” For now it is a hosted API preview available through Mistral Studio. Downloadable weights, which are what make a model genuinely open, are promised by the end of October after additional safety work. One outlet cites October 27 as the date, but Mistral itself has only said “by the end of the month.” Until the weights ship, Large 4 is a hosted product with an open-weight promise attached.
What Mistral Says It Built
Mistral describes a hybrid instruct-and-reasoning model with roughly one trillion total parameters. The active-parameter count, meaning how many parameters do the work on any single request, is reported inconsistently: early coverage said 49 billion, while Mistral’s documentation reportedly lists 1.05 trillion total and 52 billion active. Treat the active figure as unsettled. The model accepts text and images, supports more than 160 languages including every official EU language, and has a context window of one million tokens, according to Mistral’s documentation.
The training story is deliberately European. Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers and serves it from the same infrastructure. It describes the model as the first milestone funded by its €3 billion Series D, which closed in September at a post-money valuation above €21 billion. That framing is the strategic point: this is a company arguing that a frontier-scale model can be built and run outside the American and Chinese stacks.
Claims Versus Evidence
Mistral’s benchmark claims are specific, and they are its own. It reports 61.7 percent on DeepSWE v1.1, a coding-agent benchmark, and says it solved 93 percent of 40 Cybench security challenges. It reports 82 percent on a test requiring a real open-source vulnerability to be reproduced and patched, noting that two competing models refused that test. On a blind human-rated code-quality evaluation, Mistral reports 3.74, second of five models, behind Claude Opus 5 at 4.22. It also claims leadership among open-weight models on Harvey’s legal-agent benchmark.
Independent scrutiny so far is thin. A skeptical review notes that the live DeepSWE leaderboard puts Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash at about 74 percent, well above Mistral’s 61.7, and that the blind coding test was run by Mistral rather than a neutral party. The same review says claims that Large 4 beats the leading closed models are not supported even by Mistral’s own numbers, and that a broader aggregate index ranked it well down a field of dozens. Mistral itself says its reinforcement-learning run is still going and that it will publish architecture details and more benchmarks with the weights.
The fair reading is that Large 4 looks competitive for an open-weight model, particularly in security and legal work, and probably not at the top of the overall field. That would still be significant if the weights arrive with permissive terms. The license is the largest unknown: Mistral has not published final terms, and one report says the release will use a custom license rather than the Apache 2.0 license used for Large 3. A custom license could restrict commercial use, redistribution or fine-tuning in ways that matter more to startups than any benchmark.
Price and the Cost of Self-Hosting
The preview API is priced at $1.36 per million input tokens and $4.18 per million output tokens, with a launch discount reportedly cutting both roughly in half. One report compares that to Gemini 3.6 Flash at $1.50 and $7.50. The output price is the notable gap. But a trillion-parameter model is not something a small team runs on a few rented GPUs. Self-hosting requires substantial GPU capacity and shifts moderation and accountability decisions to whoever deploys it. For most startups, the realistic uses of the open weights are hosting through a cloud provider, or using Mistral’s own API, rather than running the full model in-house.
There is also a safety wrinkle. Mistral says access for cybersecurity leaders, vetted partners and state authorities will come with reduced moderation and expanded cyber capabilities, and that red-team testing is continuing with selected partners. That staged approach is more cautious than a plain weight dump, and it signals that cyber capability is where the model is strongest and where Mistral expects scrutiny.
What to Watch Next
Three things will settle the story. First, whether the weights actually land by the end of October. Second, the license text. Third, independent evaluations after release, which will show whether the self-reported cyber, coding and legal numbers hold. If they do, Large 4 becomes a credible non-American, non-Chinese option for organizations that care where their model comes from. If they do not, it is a competitive but unremarkable entry with good marketing.
What This Means for Philippine Founders
The practical value for a Philippine startup is leverage on price and vendor risk, not a new default model. Output tokens dominate costs for products that generate long text, code or reports, so a model priced near $4 per million output tokens is worth testing against whatever you use now, once independent numbers exist. An open-weight option also lets a founder keep a second supplier on hand, which matters when a single provider changes terms, pricing or availability.
Founders serving banks, insurers, government or health clients, where keeping data in a controlled environment is often a contract requirement, should watch the license and the hosting options after the weights drop. The 160-plus language claim is worth checking for Filipino and regional-language quality yourself, since Mistral’s published material emphasizes European languages and the sources reviewed here do not confirm how Tagalog, Cebuano or Taglish perform. A small evaluation set built from your own customers’ messages will tell you more than any leaderboard.
Share this article