Microsoft is aggressively expanding its frontier capabilities with the release of the MAI-Base-2 MoE model and specialized coding tools. Meanwhile, Meta's Brain2Qwerty v2 demonstrates a massive leap in neural decoding technology.
Model & Product Releases
Microsoft shipped three production-ready models this week spanning streaming transcription, foundation pre-training, and code generation.
Microsoft launches MAI-Transcribe-2-Streaming, tops Artificial Analysis
Microsoft AI · 3 min read
Microsoft released MAI-Transcribe-2-Streaming, a streaming transcription model ranking #1 on Artificial Analysis for accuracy across 60 languages with 100ms partial hypotheses, alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash voice models. Engineers building conversational AI can now get real-time transcripts with low latency and high accuracy without tradeoffs, plus matching voice synthesis.
Microsoft completes MAI-Base-2, 80B active parameter MoE model
Microsoft AI · 3 min read
Microsoft finished pre-training MAI-Base-2, an MoE foundation model with 80 billion active and 2.8 trillion total parameters that outperforms all tested open-weight models on held-out evaluations. The model achieves 4-12x better FLOPs efficiency than DeepSeek-V4-Pro-Base, giving engineers a more compute-efficient base for downstream tasks.
MAI-Code-1.1-Flash ships in GitHub Copilot at 25% token efficiency gain
Microsoft AI · 1 min read
Microsoft released MAI-Code-1.1-Flash, now in GitHub Copilot, delivering 25% greater token efficiency and 25% faster streaming at one-quarter the cost of its June predecessor. The model shows 22% improvement on Terminal-Bench 2.1 for CLI tasks and 15% on .NET, with 4% higher code survival and 9% more return visits in production.
Research
New benchmarks and clinical AI systems show measurable gains in diagnostic accuracy, non-invasive brain decoding, and agent evaluation.
Microsoft's MAI-DxO diagnostic system outperforms physicians on clinical benchmarks
Microsoft AI · 3 min read
MAI-DxO achieves 1.7x higher accuracy than physicians on MIMIC-IV-ED emergency cases and over 2x accuracy on complex NEJM cases, while reducing diagnostic costs by half to two-thirds. The system reaches correct diagnoses with fewer tests and procedures.
Meta's Brain2Qwerty v2 decodes sentences from non-invasive brain recordings in real time
AI at Meta · 2 min read
Brain2Qwerty v2 achieves real-time sentence decoding from non-invasive brain recordings at accuracy levels previously requiring surgical implants. Meta releases full training code for v1 and v2, and partner BCBL releases the v1 dataset.
Terminal-Bench 3.0 launches with 74 tasks where top models score only 34%
Terminal-Bench · 3 min read
The new benchmark spans 7 domains with harder, more diverse CLI and coding tasks that avoid saturation seen in prior versions. It introduces CI/CD pipelines, semantic versioning, and result migrations for continuous benchmark evolution.
Tooling & Infrastructure
This week brings updates to training hardware stacks, local speech synthesis, and agent evaluation benchmarks.
Amazon Trainium 3 Gets Native PyTorch Training Support
SemiAnalysis · 3 min read
AWS Trainium 3 now runs native PyTorch in eager mode with torch.compile, removing the need for XLA-style xm.step() calls. Engineers can now use standard PyTorch APIs for training on Trainium, which Anthropic already uses for production inference and RL rollouts on Trainium 2 and 3.
Ollama Adds Phonon Speech Models for Text-to-Speech
Ollama · 2 min read
Ollama now supports Phonon text-to-speech models from Gradium, offering five languages and dozens of voices streaming over 100x real-time on Apple Silicon via MLX. Engineers can integrate local TTS with `ollama run phonon` and select voices via `--voice` flag, with Apache 2.0 licensing.
Terminal-Bench 4.0 Calibrates Resources, Removes Saturated Tasks
Terminal-Bench · 3 min read
Terminal-Bench 4.0 introduces calibrated 8-hour agent timeouts, fixes tasks, and removes saturated ones, reducing timeouts and measurement noise. The update reveals Sonnet 5 struggles with output token limits (21.6B tokens vs Opus 5's 6.5B), helping engineers understand model limitations in agentic coding evals.
Terminal-Bench-Science 0.1 Benchmarks Agents on Research Workflows
Terminal-Bench · 2 min read
Terminal-Bench-Science 0.1 launches with 70 expert-curated scientific research tasks across life, physical, Earth, mathematical, and engineering sciences. Claude Opus 5 achieves 30% resolution rate, giving engineers a benchmark to evaluate agent capabilities for scientific research assistance.
Also Worth Reading
Two updates on AI's real-world deployment and platform impacts.
Meta AI Models Power Pittsburgh Assistive Robotics Project
AI at Meta · 3 min read
Meta's open-source AI models are being used by the University of Pittsburgh's HERL lab in a $41.5M ARPA-H funded project called RAMMP to build robotic assistive mobility platforms for 5.5M US wheelchair users. Engineers working on assistive robotics can reference this deployment of open models in a well-funded, real-world healthcare robotics initiative.
Google AI Search Changes Disrupt Publisher Traffic
The Verge AI · 3 min read
Google's AI search features and ranking changes have reduced referral traffic to publishers who previously relied on search optimization for audience acquisition. Engineers building content-dependent products should anticipate continued volatility in organic search distribution as AI summaries replace traditional link-based results.
The Daily Model is assembled automatically each day: 12 stories from 53 of 54 sources, summarized by a language model. Always follow the link for the original.