Microsoft AI: MAI-Image-2.6 Flash

Microsoft-Decision-1 is a small model built for fast decision-making. Instead of generating text, it reads the provided content and returns a calibrated probability for each fixed answer option, so the response carries its own confidence and can determine when an application acts, defers, or asks for review. It is post-trained from Qwen3.5-9B for single-pass scoring and suited to classification, routing, prioritization, verification, workflow control, agent guardrails, and AI judging. It is not intended for open-ended generation, conversation, translation, or summarization. Weights are updated continually while the API shape stays the same.

Learn more in Microsoft's announcement: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/

MAI-Voice-2.1-Flash is a low-latency text-to-speech model from Microsoft AI, optimized for real-time responsiveness. It produces natural, expressive speech across 23 languages, with human-like intonation, rhythm, and emotional nuance. It is suited for voice agents, assistants, call centers, and other interactive applications where latency and cost matter most.

On OpenRouter, set voice to a full voice ID with the model suffix, such as "en-US-Harper:MAI-Voice-2.1-Flash". A voice's locale sets the synthesis language. Set response_format to "mp3" or "pcm" (24 kHz mono). Harper and Grant support the agent, customer-call-center, educational, and narrator speaking styles, and many locale voices add emotion styles such as excited, happy, sad, and whispering. The full list of voices is in the supported_voices field of the models API. See the text-to-speech guide.

MAI-Voice-2.1 is Microsoft AI's highest-fidelity, most expressive text-to-speech model. It produces natural, studio-grade speech across 23 languages, with detailed prosody, nuanced expressiveness, and speaker consistency over long-form content. It is suited for audiobooks, podcasts, lectures, narration, and brand audio where maximum voice quality matters. The model prioritizes naturalness and expressivity over latency-critical generation.

On OpenRouter, set voice to a full voice ID with the model suffix, such as "en-US-Harper:MAI-Voice-2.1". A voice's locale sets the synthesis language. Set response_format to "mp3" or "pcm" (24 kHz mono). Harper and Grant support the agent, customer-call-center, educational, and narrator speaking styles, and many locale voices add emotion styles such as excited, happy, sad, and whispering. The full list of voices is in the supported_voices field of the models API. See the text-to-speech guide.

MAI-Image-2.6 is an image generation and editing model from Microsoft AI, the precision tier of the MAI-Image-2.6 family alongside the faster MAI-Image-2.6 Flash. It is suited for design-ready visuals and controlled, iterative editing, and is particularly strong at multi-reference editing that combines people, products, styles, and scenes from up to five input images while preserving composition, style, and detail. Aspect ratio can be fixed or left to the model to choose for the composition.

MAI-Image-2.6 Flash is the lower-latency, lower-cost member of the MAI-Image-2.6 family from Microsoft AI, built for latency-sensitive, high-throughput production image generation and editing at comparable quality to the precision tier. It supports the same multi-reference editing with up to five input images and the same fixed or model-chosen aspect ratios.

MAI-Transcribe 2 is a multilingual speech-to-text model from Microsoft AI, ranked #1 on the FLEURS multilingual benchmark. It supports 60 languages with automatic language identification, code switching for mixed-language speech, speaker diarization, word-level timestamps, keyword biasing for domain-specific terminology, and configurable verbatim or clean transcription styles. It is suited for captions, call transcription, subtitling, accessibility, and other voice-enabled applications, and is faster than MAI-Transcribe-1.5 on long-form audio.

On OpenRouter, set response_format to "verbose_json" for segment timestamps, and add timestamp_granularities: ["word"] for word-level timestamps. Set provider.options.azure.diarization.enabled to true for speaker labels, provide keywords through provider.options.azure.phraseList.phrases, and select a transcription style through provider.options.azure.enhancedMode.modelOptions.transcribeStyle. See the speech-to-text guide.

Microsoft AI's MAI-Image-2.5 is a high-quality image generation model available via Azure AI Foundry. It produces photorealistic and artistic images from text prompts with support for various aspect ratios.

MAI-Voice-2-Flash is a low-latency text-to-speech model from Microsoft AI for voice agents, assistants, call centers, accessibility, narration, and other interactive applications. It generates expressive 24 kHz mono speech across 15 languages and 18 locales, with fine-grained control over tone and delivery. Voice prompting and cloning require Microsoft AI-approved access and appropriate speaker consent.

MAI-Voice-2 is an expressive text-to-speech model from Microsoft AI. It is suited for conversational assistants, media narration, accessibility, education, and other long-form voice applications. It supports 15 languages across 18 locales, fine-grained control of tone and delivery, multi-speaker generation, and voice prompting from short audio clips without fine-tuning. The model prioritizes naturalness and expressivity over latency-critical generation.

MAI-Transcribe 1.5 is a multilingual speech-to-text model from Microsoft AI. It is suited for captions, call transcription, subtitling, accessibility, and other voice-enabled applications, with reliable transcription across 43 languages, diverse accents, and noisy real-world audio. It supports automatic language identification and keyword biasing for domain-specific terminology, and improves long-form transcription speed over MAI-Transcribe-1. Speaker diarization is not supported.

Microsoft AI's MAI-Image-2.5 is a high-quality image generation model available via Azure AI Foundry. It produces photorealistic and artistic images from text prompts with support for various aspect ratios.

Phi-4-mini-instruct is a lightweight open model built upon synthetic data and filtered publicly available websites - with a focus on high-quality, reasoning dense data. The model belongs to the Phi-4 model family and supports 128K token context length. The model underwent an enhancement process, incorporating both supervised fine-tuning and direct preference optimization to support precise instruction adherence and robust safety measures.

Phi-4-reasoning-plus is an enhanced 14B parameter model from Microsoft, fine-tuned from Phi-4 with additional reinforcement learning to boost accuracy on math, science, and code reasoning tasks. It uses the same dense decoder-only transformer architecture as Phi-4, but generates longer, more comprehensive outputs structured into a step-by-step reasoning trace and final answer.

While it offers improved benchmark scores over Phi-4-reasoning across tasks like AIME, OmniMath, and HumanEvalPlus, its responses are typically ~50% longer, resulting in higher latency. Designed for English-only applications, it is well-suited for structured reasoning workflows where output quality takes priority over response speed.

Phi-4-reasoning is a 14B parameter dense decoder-only transformer developed by Microsoft, fine-tuned from Phi-4 to enhance complex reasoning capabilities. It uses a combination of supervised fine-tuning on chain-of-thought traces and reinforcement learning, targeting math, science, and code reasoning tasks. With a 32k context window and high inference efficiency, it is optimized for structured responses in a two-part format: reasoning trace followed by a final solution.

The model achieves strong results on specialized benchmarks such as AIME, OmniMath, and LiveCodeBench, outperforming many larger models in structured reasoning tasks. It is released under the MIT license and intended for use in latency-constrained, English-only environments requiring reliable step-by-step logic. Recommended usage includes ChatML prompts and structured reasoning format for best results.

MAI-DS-R1 is a post-trained variant of DeepSeek-R1 developed by the Microsoft AI team to improve the model’s responsiveness on previously blocked topics while enhancing its safety profile. Built on top of DeepSeek-R1’s reasoning foundation, it integrates 110k examples from the Tulu-3 SFT dataset and 350k internally curated multilingual safety-alignment samples. The model retains strong reasoning, coding, and problem-solving capabilities, while unblocking a wide range of prompts previously restricted in R1.

MAI-DS-R1 demonstrates improved performance on harm mitigation benchmarks and maintains competitive results across general reasoning tasks. It surpasses R1-1776 in satisfaction metrics for blocked queries and reduces leakage in harmful content categories. The model is based on a transformer MoE architecture and is suitable for general-purpose use cases, excluding high-stakes domains such as legal, medical, or autonomous systems.

Phi-4 Multimodal Instruct is a versatile 5.6B parameter foundation model that combines advanced reasoning and instruction-following capabilities across both text and visual inputs, providing accurate text outputs. The unified architecture enables efficient, low-latency inference, suitable for edge and mobile deployments. Phi-4 Multimodal Instruct supports text inputs in multiple languages including Arabic, Chinese, English, French, German, Japanese, Spanish, and more, with visual input optimized primarily for English. It delivers impressive performance on multimodal tasks involving mathematical, scientific, and document reasoning, providing developers and enterprises a powerful yet compact model for sophisticated interactive applications. For more information, see the Phi-4 Multimodal blog post.

Microsoft Research Phi-4 is designed to perform well in complex reasoning tasks and can operate efficiently in situations with limited memory or where quick responses are needed.

At 14 billion parameters, it was trained on a mix of high-quality synthetic datasets, data from curated websites, and academic materials. It has undergone careful improvement to follow instructions accurately and maintain strong safety standards. It works best with English language inputs.

For more information, please see Phi-4 Technical Report

Phi-3.5 models are lightweight, state-of-the-art open models. These models were trained with Phi-3 datasets that include both synthetic data and the filtered, publicly available websites data, with a focus on high quality and reasoning-dense properties. Phi-3.5 Mini uses 3.8B parameters, and is a dense decoder-only transformer model using the same tokenizer as Phi-3 Mini.

The models underwent a rigorous enhancement process, incorporating both supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures. When assessed against benchmarks that test common sense, language understanding, math, code, long context and logical reasoning, Phi-3.5 models showcased robust and state-of-the-art performance among models with less than 13 billion parameters.

Phi-3 4K Medium is a powerful 14-billion parameter model designed for advanced language understanding, reasoning, and instruction following. Optimized through supervised fine-tuning and preference adjustments, it excels in tasks involving common sense, mathematics, logical reasoning, and code processing.

At time of release, Phi-3 Medium demonstrated state-of-the-art performance among lightweight models. In the MMLU-Pro eval, the model even comes close to a Llama3 70B level of performance.

For 128k context length, try Phi-3 Medium 128K.

Phi-3 Mini is a powerful 3.8B parameter model designed for advanced language understanding, reasoning, and instruction following. Optimized through supervised fine-tuning and preference adjustments, it excels in tasks involving common sense, mathematics, logical reasoning, and code processing.

At time of release, Phi-3 Medium demonstrated state-of-the-art performance among lightweight models. This model is static, trained on an offline dataset with an October 2023 cutoff date.