Prototype · closed beta
unOpen.AI
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Model families
Mistral.
European open weights.
For European clients, Mistral frequently resolves a conversation that has nothing to do with benchmarks: the provider is EU-domiciled, several releases carry a standard permissive licence, and the weights can run inside your own building.
General model guidance last reviewed 4 September 2026 · Timestral release update: 15 October 2026 · Engineering guidance, not legal advice.
Introduction transcript. Qubi: For many European teams, the first question is where the documents travel. Licences vary across the family, so we read each card and write down the choice. Then the model runs under your own roof, and the window stays shut. Answers come back in each language, and we still design privacy in properly.
Current release
Mistral today.
Mistral AI ships these models as a range rather than one lead model, so this page is about the family: what it is good at, where it runs and how we evaluate it.
Last reviewed on 1 September 2026. We check the public model catalogues every week; when a new release appears, we review it against the provider's announcement and update this page.
LATEST MODELS
The newest Mistral models.
What Mistral AI has shipped most recently. Each has its own page with specifications, independent ratings at every reasoning setting and listed prices.
29 Oct 2026 · open weights
Mistral Large 4
- Intelligence
- 26.0
- Context
- 1.05M
- In / out per M
- $0.700 / $2.10
29 Apr 2026
Mistral Medium 3.5
- Intelligence
- —
- Context
- 262k
- In / out per M
- $1.50 / $7.50
16 Mar 2026 · open weights
Mistral Small 4
- Intelligence
- 4.3
- Context
- 262k
- In / out per M
- $0.150 / $0.600
9 Dec 2025 · open weights
Devstral 2 2512
- Intelligence
- —
- Context
- 262k
- In / out per M
- $0.400 / $2.00
EVERY MODEL
All 20 Mistral models, compared.
Sort by any column. Select a model for its full page.
| Mistral Large 4 open | 29 Oct 2026 | 26.0 | 169 | 1.05M | $0.700 | $2.10 |
|---|---|---|---|---|---|---|
| Mistral Medium 3.5 | 29 Apr 2026 | — | 152 | 262k | $1.50 | $7.50 |
| Mistral Small 4 open | 16 Mar 2026 | 4.3 | 170 | 262k | $0.150 | $0.600 |
| Devstral 2 2512 open | 9 Dec 2025 | — | — | 262k | $0.400 | $2.00 |
| Ministral 3 14B 2512 open | 2 Dec 2025 | — | — | 262k | $0.200 | $0.200 |
| Ministral 3 3B 2512 open | 2 Dec 2025 | — | — | 131k | $0.100 | $0.100 |
| Ministral 3 8B 2512 open | 2 Dec 2025 | — | — | 262k | $0.150 | $0.150 |
| Mistral Large 3 2512 | 1 Dec 2025 | — | — | 262k | $0.500 | $1.50 |
| Voxtral Small 24B 2507 open | 30 Oct 2025 | — | — | 33k | $0.100 | $0.300 |
| Mistral Medium 3.1 | 12 Aug 2025 | — | — | 131k | $0.400 | $2.00 |
| Codestral 2508 | 1 Aug 2025 | — | — | 256k | $0.300 | $0.900 |
| Mistral Small 3.2 24B open | 20 Jun 2025 | — | — | 256k | $0.075 | $0.200 |
| Mistral Medium 3 | 7 May 2025 | — | — | 131k | $0.400 | $2.00 |
| Mistral Small 3.1 24B open | 17 Mar 2025 | — | — | 128k | $0.351 | $0.555 |
| Saba | 17 Feb 2025 | — | — | 33k | $0.200 | $0.600 |
| Mistral Small 3 open | 30 Jan 2025 | — | — | 33k | $0.050 | $0.080 |
| Mistral Large 2407 | 24 Jul 2024 | — | — | 131k | $2.00 | $6.00 |
| Mistral Nemo open | 19 Jul 2024 | — | — | 131k | $0.029 | $0.030 |
| Mixtral 8x22B Instruct open | 17 Apr 2024 | — | — | 66k | $2.00 | $6.00 |
| Mistral Large | 26 Feb 2024 | — | — | 128k | $2.00 | $6.00 |
Intelligence and coding: Artificial Analysis indices at the model's best reasoning setting, and its measured output speed. Context and prices: the OpenRouter catalogue list price in USD per million tokens. Released: the provider's date where Artificial Analysis records one, otherwise the date the catalogue first listed it. Refreshed weekly; not measured by QQuantum.ai. Select a heading to sort.
Specification
- Provider
- Mistral AI, Paris, France
- Weights
- Published for a substantial part of the range; some commercial models are API-only
- Licence
- Varies by model — several are Apache-2.0, others carry a Mistral research or commercial licence. Check the specific model card; the family is not uniform.
- Sizes
- From small models suited to a single accelerator up to large mixture-of-experts releases
- Access paths
- Mistral's own API (EU-hosted at the time of writing — confirm against the provider's current terms), Azure and other cloud catalogues, or self-hosted
- Forecasting
- Timestral, released 15 October 2026: zero-shot time-series models from 5M to 209M parameters, CPU inference, Public Preview
- Where we use it
- EU data-residency deployments, and wherever an Apache-2.0 licence is a procurement requirement
Provider-published specifications as of 1 September 2026 for general model guidance; the Forecasting row reflects the Timestral release of 15 October 2026, not figures we measured. Model families move quickly — check the provider's own documentation before committing to a number.
Where it fits well
01
EU jurisdiction end to end
A French provider with EU-hosted inference means the international-transfer analysis that dominates so many European AI projects is usually far simpler — confirm the provider's sub-processor list with your DPO. That is not a capability argument, and it decides more deployments than capability does.
02
Apache-2.0 where it matters
Several releases carry a standard permissive licence rather than a bespoke one. For organisations whose open-source policy is written around recognised licences, that difference usually means a materially shorter review.
03
Efficient sizes
The mid-size models are strong relative to their footprint, which makes them a natural fit for on-premise deployments where accelerator memory is the binding constraint and a frontier-scale model is simply not going to fit.
04
Multilingual European coverage
Solid performance across the major European languages, which matters for any organisation whose documents are not all in English and whose evaluation set should reflect that.
The honest caveats
On the hardest reasoning and coding tasks, the largest frontier models still lead. Where a workload genuinely sits at that ceiling — complex multi-step agentic work, difficult code generation — you will feel the difference, and choosing Mistral there for jurisdictional reasons is a trade you should make consciously rather than discover in production.
The licence position is not uniform across the family. Assuming everything is Apache-2.0 because one well-known release is has caught out teams before. Read the model card for the specific model you intend to deploy, and record which one you chose, because that record is what a future audit will ask for.
The family now includes forecasting models. On 15 October 2026 Mistral released Timestral, four time-series models from 5M to 209M parameters. Per the GitHub repository, the code is Apache-2.0 but the weights carry a Modified MIT licence. That weights licence grants no rights where the company or employer had global consolidated monthly revenue above US$20 million in the preceding month; a separate commercial licence can be requested from Mistral but is not guaranteed. The restriction covers the weights, not the Apache-2.0 code, and Mistral's documentation labels it Public Preview.
And EU domicile is not the same as compliance. It removes the transfer question; it does not remove your obligations around lawful basis, purpose limitation, retention or subject rights. Those still need designing into the system. See GDPR and EU data residency.
INDEPENDENT RATINGS
How Mistral compares.
Third-party scores and listed prices, refreshed every week. Useful for building a shortlist; your own evaluation decides.
Composite index across their reasoning, knowledge, maths and coding evaluations · Higher is better · Linear scale from 0 to 66.0
- Gemini 4 Carbon
- 59.4
- Claude Opus 5.5
- 58.4
- GPT-6 Astra
- 56.9
- Muse Wonder Full
- 49.7
- Mistral Large 4
- 26.0
- Mistral Large 4 Preview
- 25.3
- Mistral Small 4
- 4.3
This family's highest-rated models, highlighted, against the best-rated model of each other family we cover. One composite score cannot show cost, latency, data residency or how a model handles your task.
Source: Artificial Analysis. Checked 30 October 2026. Measured independently by Artificial Analysis, not by QQuantum.ai or the providers.
USD per million output tokens · The newest releases in the family · Linear scale from 0 to $9.00
- Mistral Large 4
- $2.10
- Mistral Medium 3.5
- $7.50
- Mistral Small 4
- $0.60
- Devstral 2 2512
- $2.00
- Ministral 3 14B 2512
- $0.20
Input is priced separately: Mistral Large 4 $0.70, Mistral Medium 3.5 $1.50, Mistral Small 4 $0.15, Devstral 2 2512 $0.40, Ministral 3 14B 2512 $0.20 per million tokens. Catalogue list prices; the provider's own rates, caching and batch discounts can differ.
Source: OpenRouter model catalogue. Checked 30 October 2026. Listed by OpenRouter, not measured by QQuantum.ai.
Questions
Questions about Mistral: European Open Weights.
Does using a French provider make us GDPR compliant?
No. It removes the international transfer question, which is often the hardest single obstacle, but lawful basis, purpose limitation, retention, subject access and the record of processing all remain yours to design. A model provider cannot be compliant on your behalf, and any vendor implying otherwise is describing a product that does not exist.
Which Mistral models are Apache-2.0?
It varies by release and it changes, so we check the model card at selection time rather than working from memory — and we write the answer into the architecture decision record, because “which licence did we deploy under” is a question that gets asked eighteen months later by someone who was not in the room.
Mistral or Llama for an on-premise deployment?
If the licence or the EU-provider question is live, Mistral. If tooling depth and the widest possible body of operational experience matter more, Llama. On raw capability at comparable sizes they trade places release by release, which is a good reason to build the system so the model is swappable.
What is Timestral, and does it change the on-premise case?
Timestral is the family of zero-shot time-series forecasting models Mistral announced on 15 October 2026. Mistral says it needs no GPU at inference and can run on client hardware, including air-gapped environments, and its GitHub repository puts the fp32 weights at 21 MB to 837 MB. The first-place benchmark rankings are Mistral's own, the technical report is still to follow and most training data was synthetic, so we would backtest it against the forecasts already in use first.
