As AI consultants based in Rosenheim, Germany, we recommend Microsoft Phi for enterprises in the DACH region (Germany, Austria, Switzerland) that want to run powerful AI on resource-constrained hardware. Since the release of Phi-4 in December 2024, the Phi models have established themselves as reliable solutions for edge and on-premise deployments, offering impressive performance with minimal requirements. With the MAI family (Build 2026), Microsoft now extends its in-house model portfolio with fully proprietary reasoning and coding models.
New: MAI-Transcribe-2 – Speech-to-Text in 60 Languages (3 September 2026)
On 3 September 2026 Microsoft AI introduced MAI-Transcribe-2 and made it available as a public preview in Microsoft Foundry, the MAI Playground and via OpenRouter. The model transcribes 60 languages – including German – and per Microsoft ranks first on the FLEURS benchmark at an average 5.2% word error rate; on the independent Artificial Analysis WER leaderboard it ranks second according to Microsoft. 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-time offer until the end of 2026 (MAI-Transcribe-1: USD 0.36 per hour) – diarization, word timestamps and keyword biasing are included.
Features: speaker diarization, word-level timestamps, keyword biasing via a phrase list, automatic language identification, code switching within an utterance, noise robustness and a selectable transcription style (verbatim with fillers for compliance and QA, clean for minutes and captions). 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. MAI-Transcribe-1.5 (43 languages, Build 2026) remains available, MAI-Transcribe-1 has been deprecated since 20 August 2026.
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 for this model. Because it is a preview without SLA, we recommend MAI-Transcribe-2 for evaluations and non-critical workloads for now; production use should wait for the GA announcement.
The MAI Family at a Glance (as of 3 September 2026)
| Model | Type | Status | EU regions per Microsoft Learn | Note |
|---|---|---|---|---|
| MAI-Transcribe-2 | Speech-to-text | Preview, 3 Sep 2026 | North Europe | 60 languages, USD 0.10/h until end of 2026 |
| MAI-Transcribe-1.5 | Speech-to-text | since 2 June 2026 | North Europe | 43 languages |
| MAI-Voice-2 / -2-Flash | Text-to-speech | Preview, 2 June 2026 | France Central, Sweden Central, West Europe | 15 languages, German voices Klaus and Mia |
| MAI-Image-2.5 / -Flash / -Pro | Image generation | Preview, 2/19 June 2026 | Sweden Central, West Europe (Global Standard) | English prompts only |
| MAI-Thinking-1 | Reasoning LLM | Preview, 2 June 2026 | all EU regions (Global Standard) | 256k context, 64k output |
| MAI-Code-1-Flash | Coding model | since 2 June 2026 | GitHub Copilot, Foundry, OpenRouter | 5B parameters |
For the GDPR assessment the deployment type is decisive: the speech models (Transcribe, Voice) process regionally in the chosen Speech region, whereas the Foundry models MAI-Thinking-1 and MAI-Image-2.5 run only as Global Standard – inference may therefore take place in any Azure region, and Microsoft does not yet offer an EU data zone for the MAI models.
MAI-Thinking-1 and MAI-Code-1-Flash (June 2026)
At Build 2026 (June 2, 2026), Microsoft unveiled its first fully in-house foundation models – trained without OpenAI distillation and exclusively on commercially licensed data. From a DACH perspective, two are especially relevant:
- MAI-Thinking-1: Sparse Mixture-of-Experts with 35B active parameters and a 256k context window (64k output). Per Microsoft it scores 97.0% on AIME 2025, 94.5% on AIME 2026 and 52.8% on 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 – exclusively as Global Standard (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-Flash: 5B-parameter coding model that per Microsoft outperforms Claude Haiku 4.5 on all four tested coding benchmarks (51.2% vs. 35.2% on SWE-Bench Pro). Rolling out in GitHub Copilot (Free, Pro, Pro+, Max) for VS Code from June 2, 2026, additionally via OpenRouter, Fireworks and Baseten.
Both models are part of the MAI family (see table above) and significantly reduce Microsoft’s dependence on OpenAI. They are, however, not open source – for full data control, Phi remains the better choice.
New: Phi-4-reasoning-vision (March 2026)
Microsoft has released Phi-4-reasoning-vision-15B, a new model that combines vision and reasoning in a 15B-parameter model. The model can autonomously decide when deeper reasoning is required — a key capability for efficient local deployments.
- 15B Parameters: More capacity than Phi-4 (14B)
- Vision + Reasoning: Understands images and can draw complex conclusions
- MIT License: Full commercial use permitted
- Self-Hosting: Runs on RTX 4090 or M2 Mac
Small Language Models (SLMs)
Microsoft Phi is Microsoft’s family of “Small Language Models” - compact but powerful models specifically optimized for efficiency.
Why Phi for Enterprises?
- Compact: 3.8B to 14B parameters
- Efficient: Runs on consumer hardware
- MIT License: Full commercial use permitted
- Self-Hosting: Full data control
- Edge-Ready: Smartphones, IoT, embedded
Key Strengths
Efficiency
Phi models achieve impressive performance at minimal size:
| Model | Parameters | Comparable Performance to |
|---|---|---|
| Phi-4 | 14B | GPT-4 (partially) |
| Phi-3.5-MoE | 42B (16 active) | Llama 3 70B |
| Phi-3.5-mini | 3.8B | Llama 3 8B |
Local Deployment
Phi models can be operated completely locally:
- Ollama:
ollama run phi4 - LM Studio: Simple GUI
- vLLM: Production deployment
- ONNX: Optimized inference
Hardware Requirements
| 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 |
Reasoning Strength
Phi-4 shows particularly strong reasoning capabilities:
- Mathematics and logic
- Coding tasks
- Structured analysis
- Chain-of-thought
Comparison to Other SLMs
| 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 |
| License | MIT | Community | Apache 2.0 |
Integration with CompanyGPT
Microsoft Phi can be integrated in CompanyGPT as a self-hosted option - ideal for enterprises that want to operate AI without cloud dependency.
Our Recommendation
For transcription in the Microsoft stack, MAI-Transcribe-2 has been the first choice for evaluations since 3 September 2026: 60 languages including German, diarization and timestamps included in the base price of USD 0.10 per hour, EU processing via a Speech resource in North Europe. Because the model is still a public preview without SLA, we plan it for production workloads only once it reaches GA; until then GPT-Realtime-Whisper and Whisper or ElevenLabs Scribe remain the production alternatives. For German-language speech output, MAI-Voice-2 (voices Klaus and Mia) from West Europe or Sweden Central is worth testing.
For reasoning-heavy cloud workloads on Azure / Foundry, MAI-Thinking-1 has been our pick in the Microsoft stack since June 2026 – 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-Flash offers a very good price-performance ratio.
For local and edge deployments with vision and reasoning requirements, we still recommend Microsoft Phi-4-reasoning-vision-15B. For pure text tasks, Phi-4 remains an excellent choice. For smartphones and IoT, Phi-3.5-mini is ideal, for simpler multimodal applications Phi-3.5-vision.
For applications requiring maximum quality where cloud hosting is acceptable, we recommend OpenAI GPT or Anthropic Claude instead.