Inside: Mira Murati's real-time AI → Gemini-powered cursor → 300M beats 27B → world's largest medical LLM drops open-source

1. Mira Murati's Thinking Machines Lab Introduces Interaction Models — Thinking Machines Lab has unveiled Interaction Models, a new class of AI architecture trained from scratch for real-time, continuous human-AI collaboration — not turn-based conversation. The system handles audio, video, and text natively and simultaneously, without the latency gaps of current voice or multimodal models. Instead of waiting for you to finish, it collaborates with you as you go. This is the first concrete output from Mira Murati's post-OpenAI lab — and it's a direct architectural challenge to how every major AI system today is built.

2. Google DeepMind Introduces an AI-Enabled Mouse Pointer Powered by Gemini — Google DeepMind has released experimental demos of an AI-powered pointer that integrates Gemini directly into the cursor. The system captures both visual and semantic context around wherever your mouse is on screen — understanding not just what's there, but what it means. It supports screen-level reasoning and interaction without switching apps or typing prompts. The cursor becomes the interface. A small UX change with very large implications for how humans interact with computers.

3. Voxtral: Mistral's full audio stack, built for voice agents Voxtral TTS clones any voice in 9 languages from a 3-second sample at 90ms latency, no fine-tuning required. It streams natively into your STT + LLM stack and handles arbitrarily long generations. Pair it with Voxtral Transcribe for end-to-end speech-to-speech. Available via API, Mistral Studio, and on Hugging Face under Apache 2.0. _(promoted)

4. Fastino Labs Open-Sources GLiGuard: 300M Parameters, Beats Models 90x Its Size — Fastino Labs has open-sourced GLiGuard, a 300M parameter safety moderation model that scores 87.7 average F1 across nine safety benchmarks — within 1.7 points of the best model, which is 90x its size. GLiGuard handles prompt safety, response safety, harm categorization across 14 categories, and jailbreak detection — all in one compact model. It runs up to 16x faster than decoder-based guard models while matching their accuracy. Apache 2.0. Drop it into any production pipeline today.

5. AntAngelMed: The Largest Open-Source Medical LLM — 103B Parameters, Only 6.1B Active — A team of Chinese researchers has released AntAngelMed, a 103B-parameter open-source medical language model built on a 1/32 activation-ratio MoE architecture — meaning only 6.1B parameters are active at inference time. You get frontier-scale medical knowledge capacity at a fraction of the compute cost. It is the largest open-source medical LLM released to date, built on top of Ling-flash-2.0 and available on GitHub now.

Meta & Stanford's Fast BLT-D Cuts Inference Memory Bandwidth 50% — No Tokenization — BLT-D processes raw bytes instead of tokens and cuts inference memory bandwidth by over 50% with no vocabulary bottleneck. The fastest and most memory-efficient model in its comparison class. A direct challenge to the tokenization assumption baked into every LLM running today.

Tilde Research's Aurora Fixes a Hidden Neuron Death Problem in Muon — Muon — the optimizer inside DeepSeek V4 and Kimi K2 — has been quietly killing neurons during training. Aurora, Tilde's leverage-aware fix, reduces dead neurons by 25% and improves training efficiency by up to 100x. Fully open-sourced and drop-in compatible with existing Muon pipelines.