微软发布七款MAI模型，Gemma 4 12B可本地运行

In today’s issue:

• Microsoft ships seven new MAI models
• MAI-Thinking-1 takes on Claude Sonnet
• Gemma 4 12B runs agents on a laptop
• NVIDIA opens 550B Nemotron 3 Ultra
• Anthropic warns of recursive self-improvement
• Agents’ Last Exam stumps frontier agents
• Claude Platform gets an ant CLI
• Cognition launches Devin Desktop
• Nous ships Hermes Desktop
• Codex builds iOS apps end-to-end
• ChatGPT memory learns to dream
• Multi-agent computer use beats solo CUAs
• Economy of Minds prices agent actions
• LEAP solves all 12 Putnam problems
• A harness rewrites itself for +19 SWE points

And all the top AI dev news, papers, and tools.

Top Stories

Microsoft Launches Seven In-House MAI Models

Microsoft AI unveiled a family of seven models trained from scratch, led by MAI-Thinking-1, its first reasoning model, in a bid for long-term self-sufficiency from OpenAI.

• MAI-Thinking-1: A 35B reasoning model that scores 97% on AIME and 53% on SWE-Bench Pro, with early testers preferring it side-by-side over Claude Sonnet 4.6 on overall quality.
• A full stack: The launch also ships MAI-Image-2.5 and Flash, MAI-Transcribe-1.5, MAI-Voice-2 and Flash, and MAI-Code-1-Flash for code generation.
• Clean training: Every model was trained on commercially licensed data with no distillation from third-party labs, which Microsoft frames as a hedge against legal risk for enterprise customers.
• Why it matters: Suleyman positions the release as a “hill-climbing machine,” a shared training infrastructure meant to keep Microsoft on the frontier as compute scales, and a direct shot at its biggest enterprise rival.

MAI-Thinking-1 ships with a detailed 109-page technical report.

Blog | Tech Report

Gemma 4 12B Brings Agentic Reasoning to Your Laptop

Google released Gemma 4 12B, a unified, encoder-free multimodal open model that brings agentic reasoning, vision, and native audio to consumer hardware under an Apache 2.0 license.

• Encoder-free design: Vision inputs pass through a single lightweight matrix multiplication and audio is projected directly into the same space as text tokens, dropping separate modality encoders.
• Runs locally: Fits in 16GB of VRAM or unified memory, small enough for a laptop, with support across LM Studio, Ollama, and Google AI Edge Gallery.
• Punches up: Reaches performance nearing Google’s larger 26B MoE model at less than half the memory footprint, and is the first mid-sized Gemma with native audio input.
• Community traction: The release topped Hacker News, with builders showing it running on a 10-year-old Xeon CPU.

Blog

Read more