Mistral Large 4: What Changed, What It Costs, Should You Switch?
Answer first
Mistral Large 4, announced October 6, 2026, is a roughly 1.05-trillion-parameter mixture-of-experts flagship in public preview via Mistral's API at $1.36/$4.18 per million tokens. Independent Artificial Analysis scoring (about 38) makes it the strongest model built outside the US and China on that index, but the promised open weights and final licence are not out yet.
What changed vs Mistral Large 3
- Scale: about 675B total parameters (Large 3) → about 1.05 trillion; active parameters 41B → 49–52B per token (sources differ, see below).
- Context window: about 256K → 1M tokens announced; the preview API currently exposes about 512K.
- Independent score: Artificial Analysis Intelligence Index about 9.3 (Large 3) → about 38 (Large 4) — one of the largest single-generation jumps logged on that index this year (via Tech Insider's launch analysis).
- Multimodal: a 1.6B-parameter vision encoder; text and image input, text output. No confirmed audio or video support.
- Trained from scratch in Europe: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters (Mistral announcement) — the sovereignty pitch is the product.
- Status: public preview only. Open weights promised by end of October 2026; the licence has not been named yet.
The numbers, with trust labels
| Benchmark | Mistral Large 4 | Source & trust |
|---|---|---|
| DeepSWE v1.1 | 61.7% | Mistral-reported (via Tech Insider, Oct 15, 2026) |
| Terminal-Bench 4 | 28.3% | Mistral-reported (via Tech Insider) |
| AutomationBench | 59.9% | Mistral-reported (via Tech Insider) |
| Intelligence Index | ~38 (GPT-6 Luna ~38; DeepSeek V4.1 Flash ~39) | Artificial Analysis — independent, as cited by benchmark trackers (Tech Insider; Doomers launch tracker) |
| Composite score | 73 ± 5.1, ranked 131st of 767 models | BenchLeader aggregator — third-party aggregation, methodology is theirs |
How to read this table: Mistral has not published a controlled head-to-head table against rival flagships, so the vendor rows and the independent rows measure different things. The independent signal is the Intelligence Index: ~38 puts Large 4 level with GPT-6 Luna and one point behind DeepSeek V4.1 Flash — frontier-adjacent, not frontier-leading. Tom's Hardware's framing, cited by Tech Insider, is that independent testing ranks it the strongest model built outside the United States and China.
What it costs (per 1M tokens, checked Oct 23, 2026)
| Route | Input | Output | Cached input | Note |
|---|---|---|---|---|
| Mistral API, standard | $1.36 | $4.18 | $0.14 | Preview pricing per Mistral's docs |
| Mistral API, Batch | $0.68 | $2.09 | $0.07 | Asynchronous, non-latency-sensitive jobs |
A 50% launch discount for the first two weeks was reported in pricing compilations; that window (from Oct 6) has passed, so do not budget on it.
Runs on what?
- API preview only, for now. Available through Mistral Studio and the Mistral API. There is nothing to download yet.
- When the weights land, self-hosting is datacentre work. A ~1.05T-parameter MoE needs multi-GPU server hardware; the ~49–52B active parameters set the serving cost, the total sets the memory floor.
- Context caveat: announced window is 1M tokens, but the preview API appears to cap at about 512K (an OpenRouter listing showed about 524K). Test your longest workloads against the shipped limit, not the announcement.
- European deployment: Mistral pitches an end-to-end European-operated deployment under European law — the differentiator versus US and Chinese providers.
Should you switch? By situation
- You are a European enterprise or public-sector buyer with data-residency requirements → Evaluate now. This is the workload the model was built and priced for, and no US or Chinese API answers the provenance question the same way.
- You run Mistral Large 3 in production → Test the preview seriously. The independent jump (9.3 → ~38) is large, but it is still a preview: validate on your tasks before migrating.
- You need open weights today → Wait. The weights and licence are promised for end of October, not published. Other open-weight models are downloadable now.
- You buy mainly on token price → Look elsewhere. At $1.36/$4.18 Large 4 is positioned on provenance, not cost; cheaper frontier-adjacent options exist.
FAQ
Is Mistral Large 4 open weights?
Not yet. Mistral describes it as an open-weight model and has promised downloadable weights by the end of October 2026, but as of October 23 the weights are not available and the specific licence has not been published. Until both land, treat it as an API-only preview.
How many parameters does Mistral Large 4 have?
About 1.05 trillion total. The active-parameter figure differs by source: Mistral's blog says 49 billion per token, its documentation says 52 billion (the gap is reportedly embeddings/output parameters). We flag the discrepancy rather than pick one.
Is the context window really 1 million tokens?
That is the announced figure. The preview API currently exposes roughly half of it — about 512,000 tokens in rollout coverage, about 524,000 on an OpenRouter listing. Which value the open-weight release will carry is not yet confirmed.
How good is Mistral Large 4 on independent benchmarks?
Artificial Analysis scores the preview at about 38 on its Intelligence Index — level with GPT-6 Luna, one point behind DeepSeek V4.1 Flash, and far ahead of Large 3's 9.3. Independent coverage ranks it the strongest model built outside the US and China on that index.
Where was Mistral Large 4 trained?
In Mistral's own datacenters in Europe, on 3,800 NVIDIA Grace Blackwell GPUs, according to Mistral's announcement. That end-to-end European training and hosting is the core of its sovereignty pitch to European buyers.
What we could not verify
- Active parameter count: 49B (Mistral blog) vs 52B (Mistral docs) — unexplained discrepancy
- Final licence terms, and the exact open-weights release date (one report named Oct 27 — unconfirmed by Mistral)
- Whether the reported 50% first-two-weeks launch discount still applies to any plan
- Full architecture details: expert count, routing and layer design are unpublished until the weights release
Sources: Mistral announcement and documentation (as compiled by Tech Insider's Oct 15, 2026 launch analysis and Pandromeda's spec tracker, checked Oct 23, 2026); Artificial Analysis Intelligence Index figures via Tech Insider and the Doomers launch tracker; pricing from Mistral's models-and-pricing documentation via Pandromeda and Eden AI. Vendor-reported figures are labelled as such wherever they appear on this card.