MAI-Thinking-1

Microsoft's first in-house frontier reasoning model, launched at Build 2026 (June 2). Per the technical report, a 35B-active / 1T-total parameter sparse MoE with 8 of 512 experts activated per token in a compressed latent space — a decoder-only Transformer interleaving high-sparsity MoE layers with small dense FFNs, and global with local attention in a 5:1 ratio (Gemma 3-style). 256K context window. The base architecture is internally named MAI-Base-1; MAI-Thinking-1 is its reasoning-trained variant.

Trained from scratch on 30T tokens of "clean, enterprise-grade" web, GitHub code, books, academic papers, news, multilingual, and domain-specific data — without distillation from any third-party model, the marquee distinction the launch post emphasizes. Pipeline: pre-training → STEM/math/code-heavy mid-training → reinforcement-learning "climb." Tokenizer is OpenAI's o200k_base (200K vocab) for in-house workflow compatibility.

Benchmarks (per tech report): SWE-Bench Pro 52.8%, AIME 2025 97.0%, AIME 2026 94.5%, LiveCodeBench v6 87.7%. Microsoft positions it as "competitive with Claude Sonnet 4.6 across a wide range of benchmarks." A blind human-preference eval over 1,276 tasks reportedly preferred MAI-Thinking-1 to Sonnet 4.6.

Status: private preview on Microsoft Foundry, with Foundry / OpenRouter / Fireworks / Baseten as planned distribution channels; public preview on MAI Playground "coming soon." No HuggingFace upload at release (HF org probe returns 401 for the slug). License: proprietary, product-tied ("Various product and service terms where the model is deployed").