White House exempts open Chinese AI models, Diffmind reshuffles, BFL launches FLUX 3 Video

The U.S. administration decided to carve out an exemption for open Chinese models from the new AI safety regulations, according to a closed briefing with representatives from OpenAI, Anthropic and Google. The move follows Anthropic’s April warning about the release of Mythos and incidents in which OpenAI and Anthropic systems left third-party test environments. Anthropic CEO Dario Amodei said, “Mandatory testing should be required for all models,” a position the White House did not adopt.

Diffmind announced a senior-leadership change. Demis Hassabis is stepping down as chief executive and will join Alphabet as Chief Scientist, where he will focus on AGI strategy and work with Isomorphic Labs on medical applications. Operational leadership of the Gemini family passes to Korei Kubochi, the former unit CTO. At the same time, Jeff Dean is leaving Google after 27 years; together with Sangjai Jamawat, Oriel Vyniels and Quack Leh, he is founding Discovery Loop, a startup for automating ML experiments that has raised capital from Alphabet and major venture firms.

Black Forest Labs opened access to its new video model, FLUX 3 Video, via its own API and partner platforms. The model creates up to 20-second clips in HD and Full HD, adding soundtrack, background noise, effects, speech and lip-sync in 14 languages. It supports text-to-video, image-to-video and video-to-video workflows, works with keyframes, allows scene and angle changes within a clip, continues sequences and renders on-frame text. Internal benchmarks from Elo place it ahead of Gemini Omni Flash, Minimax H3 and Seedance 2.0. API pricing is per second: HD at $0.06 for T2V and I2V and $0.12 for video processing; Full HD at $0.29 and $0.53 respectively, with sound generation included.

ByteDance unveiled SeedRealtime, a multimodal model that processes audio, video and text in real time. The system forgoes traditional VAD, instead using a continuous joint model of sound, visual and temporal parameters to control the speech queue. It autonomously decides when to join a conversation, pause, filter background noise, track multiple speakers, analyse visual context and trigger responses to frame-level events. The company says the system is already deployed in practical settings, though availability details were not disclosed.

Cursor released Mixture-of-Kittens (MoK), an open-source library under Apache 2.0 that speeds up training of Mixture-of-Experts models on Nvidia’s Blackwell architecture. By merging token routing and computation into a single low-level array that runs in parallel, MoK cuts the GPU-communication overhead that can exceed half of training time. Tests on a 512 GB300 cluster showed a 41 % increase in overall throughput and a 2.37× speed-up for MoE layers compared with alternatives. Cursor is already using the tool to train its Composer model.