Microsoft pursues two model lines: the Phi family of compact small language models (SLMs) for edge, on-premises and local deployments, and the MAI family (Microsoft AI) of fully in-house models for reasoning, coding, speech and images, delivered through Microsoft Foundry. Phi models are released under the MIT licence and can be fully self-hosted; the MAI models are proprietary and currently run as Global Standard in Foundry.
The current reference models are Phi-4-reasoning-vision-15B and Phi-4 for local deployments, and MAI-Transcribe-2 (including the streaming variant), MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Voice-2.1 and MAI-Image-2.6 in the cloud. As an AI consultancy from Rosenheim we classify them from a GDPR perspective for enterprises in the DACH region. This overview reflects the state as of October 2026.
Current models at a glance
MAI family
| Model | Type | Release / status | EU regions per Microsoft Learn | Note |
|---|---|---|---|---|
| MAI-Transcribe-2 | Speech-to-text | Preview, 3 September 2026 | North Europe | 60 languages, USD 0.10/h (limited until end of 2026) |
| MAI-Transcribe-1.5 | Speech-to-text | since 2 June 2026 | North Europe | 43 languages |
| MAI-Transcribe-2-Streaming | Speech-to-text (real time) | Preview, October 2026 | Sweden Central | German supported, SDK and Realtime API |
| MAI-Voice-2.1 / -2.1-Flash | Text-to-speech | Preview, October 2026 | France Central, North Europe, Sweden Central, West Europe | 23 languages, German voices Klaus, Mia, Grant and Harper |
| MAI-Image-2.6 / -2.6-Flash | Image generation | Preview, 31 July 2026 | Sweden Central, West Europe (Global Standard) | Successor to 2.5, prompts in English only |
| MAI-Image-2.7 | Image editing | Preview, 27 October 2026 | not in Foundry (Arena, Artificial Analysis, MAI Playground) | Text-to-image announced |
| MAI-Thinking-1 | Reasoning LLM | Preview, 2 June 2026 | all EU regions (Global Standard) | 256k context, 64k output |
| MAI-Code-1.1-Flash | Coding model | GA in GitHub Copilot since 11 August 2026 | GitHub Copilot; Foundry only as a preview for fine-tuning (Global) | Image input, 256K context |
For the GDPR assessment the deployment type is decisive: the speech models (Transcribe, Voice) process regionally in the chosen Speech region. The Foundry models MAI-Thinking-1 and MAI-Image, however, run only as Global Standard – inference may take place in any Azure region, and Microsoft does not currently offer an EU data zone for the MAI models. The official region availability table for “Foundry Models sold by Azure” confirms this: Llama and MAI models appear only in the Global Standard table, not in the Data Zone Standard table (which lists GPT, DeepSeek, FLUX, Mistral, Cohere and Grok models).
MAI-Transcribe-2
MAI-Transcribe-2 (3 September 2026, public preview in Microsoft Foundry, the MAI Playground and via OpenRouter) transcribes 60 languages, including German. Per Microsoft it reaches an average word error rate of 5.2% on the FLEURS benchmark and, by Microsoft’s account, second place on the Artificial Analysis WER leaderboard; on speed, Microsoft cites up to 10x faster than OpenAI GPT-Transcribe, 7x faster than ElevenLabs Scribe v2 and 5x faster than Gemini 3.5 Transcribe. The price is USD 0.10 per hour of audio as a limited offer until the end of 2026 (MAI-Transcribe-1: USD 0.36 per hour); diarisation, word timestamps and keyword biasing are included.
Features: speaker diarisation, word-level timestamps, keyword biasing via a phrase list, automatic language detection, code-switching within an utterance, noise robustness and a selectable transcription style (verbatim with filler words for compliance and QA, clean for minutes and subtitles). The model is called via the Fast Transcription API of Azure Speech (enhancedMode.model = "MAI-Transcribe-2", files up to 300 MB in WAV, MP3 or FLAC) or as input transcription in the Voice Live API.
EU assessment: Per the Microsoft Learn region list, the MAI-Transcribe model is available in eastus, northeurope, southeastasia and westus – in the EU therefore only via North Europe (Ireland). Azure Speech processes and stores data exclusively in the region of the Speech resource, so a Speech resource in North Europe confines processing to the EU. A deployment in Germany West Central is currently not possible. Because it is a preview without an SLA, we recommend MAI-Transcribe-2 for evaluations and non-critical workloads first.
MAI-Transcribe-2-Streaming
MAI-Transcribe-2-Streaming is the real-time variant for live audio: the application sends audio as a continuous stream and receives intermediate results that update the current transcript, plus final results that confirm each segment. Per Microsoft Learn the model is intended for call centers, voice assistants, live captioning, voice-driven interfaces and real-time note taking. It runs by default in multilingual mode with automatic language detection and supports German. It is connected through the Azure Speech SDK or the Realtime API (OpenAI-Realtime-compatible WebSocket protocol).
EU classification: Per Microsoft Learn, Azure serves the model from Sweden Central (EU), Central US, East US 2 and Southeast Asia and routes requests to these regions for each integration method. For EU workloads Sweden Central is therefore the option; the preview runs without an SLA and Microsoft does not recommend it for production workloads. Prices are on the Azure Speech pricing page.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash
MAI-Voice-2.1 and MAI-Voice-2.1-Flash are the current text-to-speech models of the MAI family (public preview) and have replaced MAI-Voice-2 and MAI-Voice-2-Flash in the Microsoft Learn documentation. Both support 23 languages, emotion and style control via SSML (mstts:express-as) and instant voice cloning from 5 to 60 seconds of reference audio, which is available only after limited-access approval and with the speaker’s consent. MAI-Voice-2.1-Flash targets real-time voice agents, assistants and IVR with low latency; MAI-Voice-2.1 targets high-quality long-form content such as audiobooks, podcasts, educational content and voice-overs with speaker consistency. For German, Microsoft Learn lists the voices de-DE-Klaus, de-DE-Mia, de-DE-Grant and de-DE-Harper, among others.
EU classification: Per the region list the MAI voices are available in France Central, North Europe, Sweden Central and West Europe, among others; Azure Speech processes data in the region of the Speech resource. A deployment in Germany West Central is not listed. Prices are on the Azure Speech pricing page; the predecessors MAI-Voice-2 and MAI-Voice-2-Flash (2 June 2026) are no longer listed in the current documentation.
MAI-Thinking-1 and MAI-Code-1.1-Flash
At Build 2026 (2 June 2026) Microsoft introduced its first fully in-house foundation models – per Microsoft trained without OpenAI distillation and exclusively on commercially licensed data. The coding model from that launch, MAI-Code-1-Flash, has since been replaced by MAI-Code-1.1-Flash.
- MAI-Thinking-1: sparse mixture-of-experts with 35B active parameters and a 256k context window (64k output). Per Microsoft 97.0% AIME 2025, 94.5% AIME 2026 and 52.8% SWE-Bench Pro. In Microsoft Foundry as a public preview via an OpenAI-compatible chat completions API (
/mai/v1) with function calling and encrypted chain of thought – Global Standard only (no PTU); per the region list deployable from all European Foundry regions including Germany West Central, but without an EU data zone. - MAI-Code-1.1-Flash: generally available in GitHub Copilot since 11 August 2026, via automatic model selection in Copilot Free and Student and also selectable manually in Pro, Pro+, Max, Business and Enterprise, among others in VS Code, Visual Studio, JetBrains IDEs and the Copilot CLI. New compared with its predecessor is image input. Per Microsoft the model improves on MAI-Code-1-Flash by 22% on Terminal-Bench 2.1 in the Copilot CLI and by 15% on .NET tasks, streams tokens 25% faster, needs 25% fewer tokens per task and costs a quarter of MAI-Code-1-Flash. The GitHub list price is USD 0.20 per 1 million input tokens and USD 1.20 for output. Microsoft has also introduced a 3-bit quantised version for local use with the full 256K context (recommended: more than 120 GB of RAM). In Microsoft Foundry, MAI-Code-1.1-Flash is listed as a preview for reinforcement fine-tuning, with training Global only and no data zone.
Neither model is open source – anyone who needs full data control stays with Phi.
MAI-Image-2.6 and MAI-Image-2.7
MAI-Image-2.6 and MAI-Image-2.6-Flash (31 July 2026, still preview) replace MAI-Image-2.5 / -2.5-Flash. Per the region availability table both can be deployed in the EU only via Sweden Central and West Europe, exclusively as Global Standard. The Pro variant of MAI-Image-2.5 remains listed in the model catalogue.
Microsoft introduced MAI-Image-2.7 on 27 October 2026 as a preview, initially for image editing only: restyling and relighting entire scenes, changing text and symbols in the image and restoring existing images. Per Microsoft the model improves on 2.6 across all 10 use cases and 7 editing actions that Artificial Analysis evaluates; text-to-image is to follow. So far it can be tried in Arena, at Artificial Analysis and in the MAI Playground. MAI-Image-2.7 is not yet listed in Microsoft Foundry, so MAI-Image-2.6 remains the available model for enterprise projects.
Phi family
Phi are Microsoft’s “small language models” – compact, efficient and released under the MIT licence for full commercial use and self-hosting (3.8B to 15B parameters, edge-capable for smartphones, IoT and embedded).
Phi-4-reasoning-vision-15B (March 2026) combines vision and reasoning in a 15B model and can decide for itself when deeper reasoning is needed. It runs on an RTX 4090 or an M2 Mac. Phi-4 (December 2024, 14B) is the text model with strengths in mathematics, logic, coding and chain of thought.
| Model | Parameters | Comparable performance to (per Microsoft) |
|---|---|---|
| Phi-4 | 14B | GPT-4 (partly) |
| Phi-3.5-MoE | 42B (6.6B active) | Llama 3 70B |
| Phi-3.5-mini | 3.8B | Llama 3 8B |
| Model | RAM/VRAM | Recommended hardware |
|---|---|---|
| Phi-4 | 16 GB | RTX 4070 / M2 Mac |
| Phi-3.5-MoE | 24 GB | RTX 4090 |
| Phi-3.5-mini | 4 GB | Laptop / smartphone |
| Phi-3.5-vision | 8 GB | RTX 3060 |
Locally, Phi models run via Ollama (ollama run phi4), LM Studio, vLLM (production) or ONNX.
| Feature | Phi-4 | Llama 3.2 3B | Gemma 2 2B |
|---|---|---|---|
| Parameters | 14B | 3B | 2B |
| Reasoning | Strong | Medium | Medium |
| Coding | Strong | Good | Good |
| Vision | Yes (3.5) | No | No |
| Licence | MIT | Community | Apache 2.0 |
EU availability and GDPR
- Phi (self-hosting): MIT licence, no data leaves your infrastructure – full GDPR control.
- MAI speech models: processing in the region of the Speech resource; EU via North Europe (Transcribe-2), Sweden Central (Transcribe-2-Streaming) or France Central, North Europe, Sweden Central and West Europe (Voice-2.1).
- MAI-Thinking-1 and MAI-Image: Global Standard only, no EU data zone commitment.
- MAI-Code-1.1-Flash: used via GitHub Copilot; no EU processing commitment is documented for the model, and Foundry fine-tuning runs as Global only.
- Phi in Microsoft Foundry: per the region table Phi-4, Phi-4-mini, Phi-4-multimodal and Phi-4-reasoning only as Global Standard (deployable from Germany West Central, among others), without a data zone; for EU processing, self-hosting remains the route.
Model lifecycle
- MAI-Transcribe-1 has been deprecated since 20 August 2026; MAI-Transcribe-1.5 (43 languages) remains available, and MAI-Transcribe-2 is the successor.
- MAI-Voice-2 / -2-Flash (2 June 2026) are no longer listed in the current Microsoft Learn documentation; MAI-Voice-2.1 and MAI-Voice-2.1-Flash are the successors.
- MAI-Image-2.5 / -2.5-Flash were superseded by MAI-Image-2.6; MAI-Image-2.7 has been running as a preview outside Foundry since 27 October 2026.
- MAI-Code-1-Flash (2 June 2026) has been retired across all GitHub Copilot experiences since 10 September 2026 per the GitHub changelog; the successor is MAI-Code-1.1-Flash (11 August 2026).
- The MAI models are mostly in preview without an SLA.
Integration with CompanyGPT
Microsoft Phi can be integrated into CompanyGPT as a self-hosted option – ideal for companies that want to run AI without cloud dependency.
Our recommendation
For transcription in the Microsoft stack, MAI-Transcribe-2 is the first choice for evaluations: 60 languages including German, diarisation and timestamps at a base price of USD 0.10 per hour, EU processing via a Speech resource in North Europe. Because the model is a public preview without an SLA, we plan it for production workloads only once it is GA; until then GPT-Realtime-Whisper and Whisper or ElevenLabs Scribe remain the production alternatives. For live transcription, MAI-Transcribe-2-Streaming in Sweden Central is worth testing, also as a preview. For German-language speech output, MAI-Voice-2.1 (voices Klaus, Mia, Grant and Harper) from West Europe, France Central or Sweden Central is worth testing; use the Flash variant for low-latency voice agents.
For reasoning-intensive cloud workloads on Azure / Foundry, MAI-Thinking-1 is our recommendation in the Microsoft stack – with the caveat that it runs only as Global Standard and therefore offers no EU processing commitment. For coding assistants in GitHub Copilot, MAI-Code-1.1-Flash offers good value for money; in Copilot Business and Enterprise its policy must be enabled first.
For local and edge deployments with vision and reasoning requirements we recommend Phi-4-reasoning-vision-15B, for pure text tasks Phi-4, for smartphones and IoT Phi-3.5-mini, and for simpler multimodal applications Phi-3.5-vision. For applications with the highest quality requirements where cloud hosting is acceptable, we recommend OpenAI GPT or Anthropic Claude instead.