#15Microsoft AI

MAI-Image-2.6

MAI-Image-2.6 is Microsoft AI's newest image generation and editing model, released with a faster MAI-Image-2.6-Flash variant. It adds editing that combines several reference images and web grounding for current information, at up to 1.5K resolution. Microsoft reports No. 2 rankings on Arena for both text-to-image and image editing.

• Text-to-image and image editing; Flash variant for speed
• Multi-reference editing and web grounding
• Up to 1.5K resolution with dynamic aspect ratios
• No. 2 on Arena text-to-image and editing (Microsoft)

AccessMAI Playground (free trial) and Microsoft Foundry (public preview); MAI-Image-2.6-Flash variant

MAI-Transcribe-2

MAI-Transcribe-2 is Microsoft AI's batch speech recognition model, adding speaker separation, word-level timestamps, keyword biasing and adjustable transcript styles. It covers 60 languages and handles noisy audio and speakers who switch languages. Microsoft reports a 5.2% average word error rate on the FLEURS benchmark and says it is faster than OpenAI, ElevenLabs and Google rivals.

• Speech-to-text in 60 languages
• Speaker diarization and word-level timestamps
• Keyword biasing and configurable transcription styles
• FLEURS average word error rate 5.2% (Microsoft)

AccessMicrosoft Foundry, MAI Playground, OpenRouter

Price$0.10 per hour of audio (limited-time offer through year-end)

MAI-Thinking-1

MAI-Thinking-1 is Microsoft AI's first flagship reasoning model, trained in-house from scratch without distillation from other companies' models. It is a sparse mixture-of-experts model with about 1 trillion total and 35 billion active parameters. Microsoft pitches it as a mid-priced model for complex enterprise work, with strong coding and maths scores.

• 256K-token context window
• ~1T total / 35B active parameters, sparse mixture-of-experts
• AIME 2025 97.0%, AIME 2026 94.5% (Microsoft)
• Matches Claude Opus 4.6 on SWE-Bench Pro, per Microsoft

AccessMicrosoft Foundry (public preview) and MAI Playground; Chat Completions-compatible API

MAI-Code-1.1-Flash

MAI-Code-1.1-Flash is Microsoft AI's lightweight agentic coding model built into GitHub Copilot and VS Code. It replaces MAI-Code-1-Flash, launched in June 2026. Microsoft says it writes higher-quality code, uses 25% fewer tokens per task and costs a quarter of the earlier version.

• Agentic coding inside GitHub Copilot and VS Code
• 22% better on Terminal-Bench 2.1 than 1.0 (Microsoft)
• 25% fewer tokens per task; 25% faster streaming
• Priced at one quarter of MAI-Code-1.0

AccessGitHub Copilot and VS Code

MAI-Voice-2

MAI-Voice-2 is Microsoft AI's second-generation text-to-speech model for production voice apps. It offers emotion tags, voice cloning from a short clip and a steady speaker identity over long content such as audiobooks. Microsoft says listeners preferred it to MAI-Voice-1 72% of the time.

• 17 languages and locales
• Emotion control via tags such as whispered or excited
• Zero-shot voice cloning from 5-60 seconds of audio
• Hindi-English and Spanish-English code-switching

AccessMicrosoft Foundry API; used in VS Code and Dynamics 365 Contact Center

Phi-4-reasoning-vision-15B

Phi-4-reasoning-vision-15B is a 15-billion-parameter open-weights vision-language model from Microsoft Research. It decides when to reason step by step and when to answer directly, and handles charts, documents, maths and science questions and finding elements on computer screens. Microsoft says it trained on about 200 billion tokens of multimodal data.

• Image and multi-image reasoning, captioning and visual Q&A
• Reads documents, receipts and charts
• Screen/UI grounding for computer-use agents
• ~75% average on ChartQA, MathVista, MMMU, ScreenSpot (Microsoft)

AccessOpen weights on Hugging Face, Microsoft Foundry and GitHub, under a permissive licence (Microsoft Research)

Fara-7B

Fara-7B is Microsoft Research's first agentic small language model for computer use. With 7 billion parameters, it looks at browser screenshots and completes web tasks by scrolling, typing and clicking, and is small enough to run on a PC. Microsoft released it as open weights under the MIT licence.

• Operates a browser from screenshots: click, type, scroll
• 7B parameters; can run locally on Copilot+ PCs
• WebVoyager 73.5%, Online-Mind2Web 34.1% (Microsoft)
• Introduced the WebTailBench benchmark

AccessOpen weights under MIT licence on Hugging Face (microsoft/Fara-7B) and Microsoft Foundry; runs on Copilot+ PCs via AI Toolkit for VS Code