Microsoft releases MAI-Thinking-1, its first reasoning AI model trained without third-party distillation
Microsoft announced MAI-Thinking-1, its first advanced reasoning AI model, at Build 2026. The company claims it's a medium-sized model matching leading models on key software engineering benchmarks, trained from scratch without distillation from third-party models.
Microsoft releases MAI-Thinking-1, its first reasoning AI model trained without third-party distillation
Microsoft announced MAI-Thinking-1, its first advanced reasoning AI model, at Build 2026 on June 2. The company describes it as a "medium-sized model" that "matches leading models" on "key" software engineering benchmarks, according to Microsoft's claims.
The model represents Microsoft's most ambitious in-house development effort since introducing its initial proprietary models in 2025. Prior to that, the company relied exclusively on OpenAI's models. Microsoft and OpenAI recently renegotiated their partnership to loosen ties between the companies.
Microsoft claims MAI-Thinking-1 was "trained from the ground up on clean data, without distillation from third-party models." Specific benchmark scores, context window size, and pricing have not been disclosed.
Six additional models announced
Microsoft unveiled six other models at Build 2026:
MAI-Image 2.5 and its flash variant handle text-to-image generation and image editing.
MAI-Transcribe-1.5 is claimed to be "five times faster than competing models," though Microsoft did not specify which models it's comparing against.
MAI-Voice-2 adds 15 new languages and expanded voice options. A flash version is listed as "coming soon."
MAI-Code-1 is described as "inference-efficient" and has been integrated into GitHub Copilot and Visual Studio Code.
Microsoft did not provide technical specifications, pricing details, or release dates for general availability of any of these models.
What this means
Microsoft's announcement signals a strategic shift toward proprietary model development after years of OpenAI dependency. The emphasis on training "without distillation" suggests Microsoft is positioning MAI-Thinking-1 as a fully independent reasoning model, though without disclosed benchmarks or third-party testing, performance claims remain unverified. The simultaneous release of seven models across multiple modalities indicates Microsoft is building a comprehensive model family to compete directly with Anthropic, OpenAI, and Google—but concrete details on capabilities and availability remain limited.
Related Articles
MiniMax releases M3.1-Flash: 428B-parameter multimodal MoE with 1M-token context and weights on Hugging Face
MiniMax has published MiniMax-M3.1-Flash on Hugging Face: a native multimodal mixture-of-experts model with about 428B total parameters, 23B activated per token, and a context window of up to 1M tokens. It builds on MiniMax-M3 with sparse attention, quantization-aware training and DSpark speculative decoding.
Xiaomi open-sources Xiaomi-Lens-0, a 4B multimodal model that outputs pixel-level mask tokens mid-answer
Xiaomi's SeerRay-Lab has open-sourced Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5-4B. The model emits mask tokens for pixel-level image regions in the middle of an answer. The lab also released a benchmark called LensBench.
Microsoft's ThinkingBox-Training lifts Qwen3.8-27B fully reliable task completions from 38 to 107 of 507
Microsoft and Hugging Face released ThinkingBox-Training, which turns executable database-state checks into RL rewards. After 50 GRPO steps, Qwen3.8-27B's pass@1 rose from 51.70% to 60.78%, and tasks passed on all 20 attempts rose from 38 to 107, according to the authors.
MOSS-Image, a 5B open-weights image model, passes one million downloads in four days
MOSS-Image, an open-weights text-to-image model from MOSS (模思智能) and the OpenMOSS team, is built on a 5B-parameter DiT backbone and has passed one million downloads in four days, according to Chinese outlet QbitAI. It supports structured JSON layout input and text prompts up to 4096 tokens.
Comments
Loading...