Mistral AI is a French AI company headquartered in Paris. As an EU-native company, Mistral offers enterprises in the DACH region with strict privacy and compliance requirements models that can be operated in the EU – via the EU endpoint of the Mistral API, via Microsoft Foundry (Data Zone Standard EU) or by self-hosting open weights. The portfolio ranges from language models (Large 4, Medium 3.5, Small 4, Large 3) through audio (Voxtral 3) and code (Codestral) to document AI (OCR 4.1) and time-series forecasting (Timestral).
For production use, the reference models are Mistral Medium 3.5 (flagship for agents, coding and reasoning), Mistral Small 4 (efficient model under Apache 2.0), Mistral Large 3 (open-weight model under Apache 2.0) and Mistral OCR 4.1 for documents. Mistral Large 4, Voxtral 3 and Timestral are in public preview. As an AI consultancy from Rosenheim we support model selection, GDPR assessment and operations, including through CompanyGPT. As of October 2026.
Current models at a glance
| Model | Status | Context | Licence | Focus |
|---|---|---|---|---|
| Mistral Large 4 | Public preview (6 October 2026) | 1M | Open weights announced, licence not yet published | Coding, agents, multimodal tasks, cybersecurity |
| Mistral Medium 3.5 | GA (29 April 2026) | 256k | Open weights (modified MIT licence) | Agents, coding, reasoning |
| Mistral Small 4 | GA (16 March 2026) | 256k | Apache 2.0 | Efficient standard workloads, self-hosting |
| Mistral Large 3 | GA (December 2025) | 256k | Apache 2.0 | Complex enterprise applications, agentic workflows |
| Voxtral 3 | Public preview (20 October 2026) | – | API (Premier) | Transcription, real-time speech, speech synthesis, voice agents |
| Mistral OCR 4.1 | GA (31 August 2026) | – | API | Document processing |
| Timestral 1.0 | Public preview (15 October 2026) | 32K time steps | Modified MIT licence | Time-series forecasting |
Mistral Large 4
Mistral Large 4 (mistral-large-4) is, per the model card, an open, multimodal mixture-of-experts model with 1.05 trillion total parameters, 52 billion active parameters, a 1.6-billion-parameter vision encoder and 1M context. It combines instruct, reasoning and agentic capabilities in a single model; per Mistral the training data covers more than 160 languages, including all official EU languages. Since 6 October 2026 it has been available as a public preview via the Mistral API. Per Mistral it was trained on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres in Europe, and the preview is served on the same infrastructure.
Mistral cites, among others, 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 93% on Cybench (per Mistral, own measurements). Mistral has announced the weights for the end of October 2026; the model card lists them as “Coming soon”. Until they are published, the licence and self-hosting requirements remain open. For Large 4 this means: evaluate yes, run production workloads only after GA and publication of the weights.
Mistral Medium 3.5
With Medium 3.5 (29 April 2026) Mistral released its first merged flagship model. It combines the previously separate models Magistral (reasoning) and Devstral 2 (agentic coding) in one dense 128B model with 256k context, a switchable reasoning mode and native image processing. According to Mistral it reaches 77.6% on SWE-Bench Verified. The weights are available under a modified MIT licence. Medium 3.5 also powers the Vibe agent (see below). Since 28 May 2026 the model is also available in Microsoft Copilot Studio with in-region processing for EU customers.
Mistral Small 4
Small 4 (16 March 2026) uses a sparse MoE architecture with 119 billion parameters, of which only about 6 billion are active (4 experts). It offers 256k context, native vision and the reasoning_effort parameter to control reasoning depth at runtime. According to Mistral, latency drops by about 40% compared with Small 3 while throughput triples. The Apache 2.0 licence allows unrestricted self-hosting.
Mistral Large 3 and specialised models
Mistral Large 3 (December 2025) has 675 billion total parameters (41 billion active), 256k context and, per the model card and the Hugging Face repository, is released under Apache 2.0. Specialised models include:
- Codestral: code generation and analysis
- Ministral 3 family: edge models in 14B, 8B and 3B for on-device use
- Voxtral Transcribe 2 and Voxtral Small (25.07): predecessors of the Voxtral 3 family, still listed in the documentation
Voxtral 3
On 20 October 2026 Mistral released the Voxtral 3 family as a public preview, together with a voice agent SDK. The four models per the Mistral changelog:
- Voxtral Small Transcribe 3 (
voxtral-small-transcribe-3-0): transcription and diarisation in 50+ languages, with context biasing and end-of-turn detection. - Voxtral Mini Transcribe 3 (
voxtral-mini-transcribe-3-0): efficient offline transcription (4B parameters) with diarisation and timestamps. - Voxtral Mini Transcribe Realtime 3 (
voxtral-mini-transcribe-realtime-3-0): real-time transcription with latency under 500 ms for agent applications. - Voxtral Mini TTS 3 (
voxtral-mini-tts-3-0): speech synthesis in 35+ languages and 50+ dialects with zero-shot voice cloning and context-dependent emotion; Pro Voice can reproduce a speaker from about 30 minutes of audio.
The voice agent SDK (built on LiveKit) joins transcription, any language model with tool calling and speech synthesis into one streaming pipeline, over SIP for telephony or WebRTC in the browser. Per Mistral, Voxtral 3 is available through the Mistral API with European, US and global endpoints; for operation on your own infrastructure Mistral refers to an enquiry with its sales team. The benchmark comparisons (word error rate, diarisation) come from Mistral.
Mistral OCR 4.1
On 31 August 2026 Mistral declared OCR 4.1 (mistral-ocr-4-1) generally available; the first release dates from 16 July 2026 per the changelog. The aliases mistral-ocr-latest and mistral-ocr-4 point to OCR 4.1 – anyone using them is already on the new model and should re-check their pipeline outputs. Per Mistral’s documentation, OCR 4.1 offers:
- Paragraph-level bounding boxes – native, without a downstream layout analysis.
- Structural block labels that type layout elements.
- Block-level confidence scores: the
confidence_scores_granularityparameter supportspage,blockandword. In invoice or contract processing, uncertain blocks can be routed to manual review. - Structured annotations and bbox extraction via
/v1/ocr, bulk processing via/v1/batch.
Timestral 1.0
Timestral (15 October 2026, public preview) is Mistral’s first family of time-series foundation models in four sizes (5M, 28M, 123M and 209M parameters). The models deliver zero-shot forecasts with 21 quantiles, handle contexts of up to 32K time steps as well as past-only and known-future covariates, and per Mistral also run on CPU. Per Mistral the 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. The weights are available under a modified MIT licence on Hugging Face (mistralai/Timestral-209M and others); Mistral names on-prem, VPC or air-gapped operation. Typical use cases are capacity planning, energy, supply chains and predictive maintenance.
Pricing
Provider prices per the Mistral model cards and price list (per 1M tokens, input / output):
| Model | Input | Output | Note |
|---|---|---|---|
| Mistral Large 4 (preview) | USD 1.36 | USD 4.18 | Cached input USD 0.14; the 50% launch discount applied for the first two weeks after launch |
| Mistral Medium 3.5 | USD 1.50 | USD 7.50 | |
| Mistral Small 4 | USD 0.15 | USD 0.60 | |
| Mistral Large 3 | USD 0.50 | USD 1.50 |
Audio (Voxtral 3, preview): Small Transcribe 3 USD 0.004 per minute, Mini Transcribe 3 in batch USD 0.0035 per minute, Mini Transcribe Realtime 3 USD 0.008 per minute, Mini TTS 3 USD 13 per 1M characters. Per Mistral, the voice agent API costs USD 0.04 per minute with Small 4 as the reasoning model and USD 0.08 per minute with a larger model.
OCR costs USD 4 per 1,000 pages, and USD 5 per 1,000 pages with annotations (Document AI). Compared with OCR 3 (around USD 2 per 1,000 pages) the per-page price has doubled – for high volumes it is worth calculating whether the new structural data justifies the premium. The EU endpoint of the Mistral API is billed at 1.1 times list prices (a 10% upcharge).
EU availability and GDPR
Mistral is a French company subject to European data protection law. Per its documentation, the Mistral API offers three endpoints: global (api.mistral.ai, no fixed processing location), EU (api.eu.mistral.ai, multiple data centres in EU and EFTA countries) and US. The EU endpoint keeps inference processing in the EU; account management, billing and other metadata may still be processed outside the region per Mistral. Agents, Batch and the Files API are not available on regional endpoints, and model availability differs by region (check with models.list).
| Model | Mistral API | Microsoft Foundry | AWS Bedrock / Google Vertex AI | Self-hosting |
|---|---|---|---|---|
| Large 4 (preview) | Preview; EU endpoint not stated | not documented | not documented | Weights announced |
| Medium 3.5 | yes; check EU endpoint in docs | Data Zone Standard EU | not listed | Open weights (modified MIT) |
| Small 4 | yes; check EU endpoint in docs | not listed | not listed | Apache 2.0 |
| Large 3 | yes; check EU endpoint in docs | Data Zone Standard EU | Bedrock: Stockholm per model card, region overview without EU region; Vertex not listed | Apache 2.0 |
| Voxtral 3 (preview) | EU, US and global endpoint per Mistral | not documented | not documented | on request |
| OCR 4.1 | yes | not listed (Foundry lists OCR 4.0 and Document AI 25.12) | Vertex: only OCR 25.05 | – |
| Timestral (preview) | – | – | – | Open weights (modified MIT) |
Details on regions and model versions at the hyperscalers are in the compliance table above. London regions are in the UK, not the EU.
Model lifecycle
Mistral maintains a table of deprecations and retirements on docs.mistral.ai. Mistral Large 3 and Medium 3.5 are not listed there as deprecated. Per that table the following are retired:
- Mistral Medium 3 (31 August 2026), Mistral Small 3.2, Magistral Medium 1.2, Devstral 2 (each 31 July 2026) and Pixtral Large (31 May 2026) – successors are Medium 3.5 and Small 4 respectively. Magistral and Devstral 2 have been merged into Medium 3.5.
- OCR 4 (
mistral-ocr-4-0, 30 September 2026): the successor is OCR 4.1 at the same price. OCR 3 remains available for existing integrations per the documentation. - Leanstral 1.5 (2 July 2026): open-weights model (119B MoE, 6B active, Apache 2.0,
mistralai/Leanstral-1.5-119B-A6B) for proof engineering in Lean 4. The Labs API endpoint (labs-leanstral-1-5) was retired on 30 September 2026; self-hosting the weights remains possible.
Other developments at Mistral
- Robostral Navigate (8 July 2026): first robotics model, 8B parameters for embodied navigation from natural-language instructions and a single RGB camera. According to Mistral 79.4% (val-seen) and 76.6% (val-unseen) on R2R-CE. Mistral has not announced licence or availability.
- Vibe: since late May 2026 Le Chat is called Vibe – an agent for work and code with Work Mode (connectors to Google Workspace, Outlook, SharePoint, Slack, GitHub) and Code Mode (remote coding, pull requests), powered by Medium 3.5.
- Funding: on 8 September 2026 Mistral announced a EUR 3 billion Series D led by Samsung Electronics at a post-money valuation of over EUR 21 billion (per Mistral).
Integration with CompanyGPT
The Mistral models that are generally available can be operated in CompanyGPT in a GDPR-compliant way in the EU – via the Mistral API with the EU endpoint, Microsoft Foundry (Data Zone Standard EU) or self-hosted; on AWS Bedrock and Google Vertex AI, EU region availability depends on the model version (see the compliance table above). We deploy preview models (Large 4, Voxtral 3, Timestral) in customer projects only after individual review.
Our recommendation
We recommend a two-track strategy: Mistral Medium 3.5 for enterprise agents, coding and reasoning – thanks to toggle reasoning, 77.6% SWE-Bench Verified (per Mistral) and EU data residency including Copilot Studio integration. Mistral Small 4 remains the cost-efficient choice for standard workloads thanks to its Apache 2.0 licence and MoE efficiency. For large-class open-weight models there is Mistral Large 3 (Apache 2.0). Until the weights are published and GA is reached, we classify Mistral Large 4 as a preview: suitable for evaluation and tests with non-critical data, not yet a production model; licence and EU operation need to be checked. For specialised workloads we name Voxtral 3 (audio, pilot projects during the preview), Mistral OCR 4.1 (documents), Timestral (time series, preview), the Ministral 3 family (edge) and Leanstral 1.5 (formal proofs, self-hosting of the Apache 2.0 weights only).