Mistral released Mistral Large 4, nicknamed “Le Chonk”, in public API preview on 6 October. It is a roughly 1-trillion-parameter mixture-of-experts model with 49B active parameters, listed at $1.36 per million input tokens and $4.18 per million output tokens, and open weights are currently expected on 31 October. It is not the best general-purpose model you can buy. It is a credible European contender on legal and finance agent work, priced well below the closed frontier tier, so for most teams the right move this month is a targeted evaluation rather than a stack migration.
What Mistral launched
According to Kingy AI’s source-checked breakdown of the Large 4 specs, benchmarks and pricing, the model has 1.05 trillion total parameters. Of those, 49B sit in active experts, or 52B if you count embeddings and output layers. It also has a 1.6B-parameter vision encoder, and Mistral’s model card lists a one-million-token context window. Independent evaluators and gateways currently list around 512K to 524K, so verify the limit on the endpoint you actually call. Training data covers more than 160 languages, including every official EU language. The model takes text and images as input and returns text.
The weights are not downloadable yet. Mistral’s Hugging Face page lists 31 October as the expected date, and the download and licence sections still read “coming soon”. Mistral VP Science Pierre Stock told TechCrunch that the weights will follow safety testing, and that Mistral will work with trusted partners and governments in the meantime. Large 3 shipped under Apache 2.0. Do not assume Large 4 will get the same licence until the terms are published.
How good is it?
The results are uneven, and that tells you more than any single headline number. On Vals AI’s broad index, Large 4 scores 48.05%, which ranks it 32nd of 44 and puts it behind GPT-6 Astra, Claude Sonnet 5.5 and Claude Opus 5.5. On the Harvey Legal Agent Benchmark it scores 15.83%, sixth of 75 and ahead of all three of those closed models. On Finance Agent v2 it edges Astra, 54.68% to 53.54%. Coding is the weak spot. On Vals’ Terminal-Bench 4 it scores 22.73%, against 65.15% for Opus, and CNBC reports that the model still lags the frontier in areas such as coding.
Mistral’s claim to beat closed models on visual grounding rests on Dense200, where it reports 42% against 41% for GPT-6 Astra. That is one point on one test. On Artificial Analysis’s GDP.pdf document-reasoning benchmark, Large 4 scores 18.6% and Astra scores 32.2%. Simon Willison notes that the model scores 38 on Artificial Analysis’s index, compared with 9 for Large 3. He puts it at maybe about six months behind the frontier. That is a big jump for Mistral.
The price, and the small print
$1.36/$4.18 are the list rates. Mistral is currently running a 50% launch discount at $0.68/$2.09. Vercel’s gateway listing says the promotion ends on 20 October, so budget on the full price.
| Model | Input /M | Output /M |
|---|---|---|
| Mistral Large 4 (list) | $1.36 | $4.18 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-6 Astra (short-input tier) | $10.00 | $50.00 |
A low token price does not guarantee a low cost per task. Vals reports $13.78 per index test for Large 4, against $3.24 for GPT-6.1 Sol. Those figures were measured at Large 4’s higher launch rates, and longer agent runs consume tokens. The useful figure is your total spend divided by the number of outputs you actually accept.
The buried story: who owns the compute
Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centres. Stock and CNBC round this to about 4,000, and CNBC says training took two months. Stock says that is two to three times fewer GPUs than Mistral’s Chinese competitors use. Only weeks ago, Mistral was pushing back on suggestions that hosting Chinese models made it a mere inference provider. Now it has a frontier-adjacent model trained on infrastructure it controls. If your planning assumes open weights always trail by a full generation, update the assumption. Willison’s estimate is closer to six months.
Why it matters for Irish, UK and EU businesses
The first benefit is procurement leverage. A credible European fallback makes single-vendor lock-in harder to justify. Mistral offers EU and US regional inference routes for a 10% price uplift. Its documentation separates inference processing from control-plane data such as accounts and billing, so check both against your data security and privacy requirements.
Self-hosting is a different proposition. At two bytes per parameter, the raw weights come to roughly 2.10 TB, and idealised four-bit packing still needs around 525 GB. This is a multi-GPU cluster project, not a laptop. For most SMEs it only pays off with sustained high utilisation or strict data constraints.
What to do now
- Test it on a narrow slice of your real workload where it looks strong, such as contract review or financial research agents. Define acceptance criteria before you look at any outputs.
- Keep a frontier generalist available for coding agents and other tasks where Large 4 trails.
- Put model calls behind an abstraction layer so that switching providers is a config change, not a rewrite. This is the pattern we describe in integrating LLMs into legacy systems.
- Wait for the published licence before you commit budget to a self-hosted build.
If your product’s margin depends on the price gap between closed and open models, Large 4 has just narrowed that gap. A moat built on that price gap is shrinking, which is why we build on systems that compound: proprietary data, evaluations and workflow integration.
Frequently Asked Questions
Can I download Mistral Large 4 today?
Not the weights. You can use the model now through Mistral's own API or through third-party gateways such as Vercel's AI Gateway. Mistral's Hugging Face page lists 31 October as the expected weight release, and the licence has not yet been published.
Is $1.36 per million input tokens the price I will pay?
That is the list rate. Mistral is currently charging a discounted $0.68 input and $2.09 output per million tokens, and Vercel's listing says the promotion ends on 20 October. EU regional inference adds 10%.
Should we replace our current model with Large 4?
Not wholesale. Its independent results are strong on legal and finance agent benchmarks and weak on broad, coding and document-reasoning indices. Evaluate it on specific workloads and keep a frontier generalist as a fallback.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
