AI Lab
Official Microsoft announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
MAI Image 2.7 debuts in preview at No. 3 on Arena Edit and No. 2 on Artificial Analysis's image editing leaderboard, above GPT Image 2.5 Flare. It improves on 2.6 across all 10 use cases and all 7 editing actions, with a 24 Elo lead in scene and style.
MAI Image 2.6 gains +79 Elo over MAI Image 2.5 on the Arena text-to-image leaderboard, with text rendering alone up +91 Elo. It is generally available in Microsoft Foundry, with MAI Playground access later this week.
MAI Cyber 1.1 Flash matches MAI Cyber 1 Flash performance while costing 18% less per scan on Microsoft's internal MDASH vulnerability-scanning suite. MAI Cyber now accounts for almost 50% of all MDASH scans, and nearly 9 in 10 follow-up scans use the MAI profile over the OpenAI-only profile.
Hugging Face post details post-training agents with ThinkingBox, using database grading outcomes as reinforcement-learning reward signals. No benchmark scores or model versions are given in the available snippet.
Flint-interactive lets agents build charts users can click, filter, zoom, highlight, and cross-link, using high-level semantic events instead of low-level visualization code. Flint's compiler handles interaction logic and visual dependencies, avoiding common layout and rendering bugs.
Researchers resolved the Kannan-Lovász-Simonovits conjecture, open since 1995, showing the best flat cut matches the best cut of any shape within a universal constant regardless of dimension. The result came from a human-directed agentic process where AI explored possibilities and researchers pruned and redirected, then machine-checked formal verification.
MAI-Base-2 is an MoE foundation model with 80B active and 2.8T total parameters, which Microsoft says outperforms every open-weight model it tested on held-out evaluations. It is 4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base; post-training is ongoing.
StudentSim trains personalized AI student simulators from public, de-identified learner records, one simulator per student, to evaluate tutors before real-learner studies. It beats leading baselines across chess, second-language English writing, and math, scoring on whether simulators respond like a specific student and update after tutor guidance.
Microsoft-Decision-1 returns a calibrated probability per fixed answer option instead of generated text, scoring 83.5% average accuracy across 36 vendor-run benchmarks at 85 ms p50 latency. It is post-trained on Alibaba's Qwen3.5-9B, not an OpenAI model, and priced at $0.042 per million input tokens with free output.
MAI-Code-1.1-Flash is 25% more token-efficient and a quarter the price of the model Microsoft launched in June at Build, with 22% better Terminal-Bench 2.1 scores in Copilot CLI and 15% gains on .NET tasks. A 3-bit quantized build runs locally with a 256K context window; Microsoft recommends 120GB+ RAM.
Agent Lightning v1.0 is a ~3,500-line framework that trains agents using the same harness they run in deployment, with agents running as standard Kubernetes jobs. An end-to-end coding pipeline lifted Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified using about 6,000 training samples.
A Hugging Face blog post titled "The Agent Said It Was Done. The Database Disagreed." examines the gap between an AI agent's self-reported task completion and actual database state. No further details were available in the source snippet.
FrogNano-4B is derived from Qwen3.5-4B and keeps its dense 32-layer hybrid Gated DeltaNet and gated-attention architecture, with text-only post-training for repository-level software engineering. It was trained via RL on ~1,500 synthetic SWE task environments using the five-tool Leaf harness, without distilling stronger-model trajectories.
MAI-Transcribe-2-Streaming hits 2.5% WER at 0.13s after end of speech, ranking #1 of 38 models on AA-WER Streaming for both final and first partial transcripts. It covers 60 languages with partials in ~100ms, priced at $0.54/hour of audio through year-end.
The pipeline detected nearly 80% of major space-weather events in evaluation and warns grid operators 30-60 minutes before a specific risk appears. It combines Auroral Electrojet and Disturbance Storm Time forecasts with local latitude, geology, and ground conductivity.