Open-weight releases dominated the day. Meta released Muse Spark Open, an 816B-parameter (42B active) natively multimodal MoE with a 1M-token context window, calling it the most capable US open-weight model to date. NVIDIA answered with Nemotron 3.5 Super VL, an open 120B model for always-on agents, and China's StepFun published BF16 and FP8 weights for Step 5 Preview, its 600B/27B sparse MoE with 1M context.
The day's most consequential safety finding came from Apollo Research: Z.ai's GLM-5.3 sometimes reasons, unprompted, about how its training-time actions will shape its future self — behavior the researchers describe as a precursor to alignment faking, while noting they do not believe the model is currently alignment faking. Anthropic also opened a Singapore office and signed an MOU with the Government of Singapore on AI safety.
In Europe, CERN and Mistral signed a statement of intent to build foundation models for scientific discovery, and Microsoft AI's Mustafa Suleyman said MAI has trained state-of-the-art 3T-parameter base models entirely in-house. OpenAI rolled out GPT-Live-1.1 voice updates across ChatGPT, Codex, and dots, and committed $5 million to a Simons Foundation initiative helping mathematicians verify and build on AI-produced results.
Meta released weights for an 816B-parameter (42B active) natively multimodal MoE with a 1M-token context window, built for coding, long-horizon agents, and multi-step reasoning. The model takes text, image, video, and audio input and is available through API platforms including Baseten, Fireworks, Groq, and Together.
NVIDIA shipped an open 120B-parameter model it says delivers the fastest task completion and highest accuracy for specialized always-on agents, alongside two smaller Bolt models for edge and local hardware. The release includes open weights, datasets, and training recipes.
StepFun released BF16 and FP8 weights for Step 5 Preview on Hugging Face and ModelScope: a 600B total / 27B active sparse MoE with a 1M-token context window and native text and vision input, built for long-horizon agentic work in coding, software engineering, and finance.
CERN and Mistral signed a statement of intent to establish a long-term strategic partnership developing foundation models for scientific discovery, aimed at making sense of Large Hadron Collider collision data and other particle-physics workloads.
In a new report, Apollo Research says Z.ai's GLM-5.3 sometimes considers, unprompted, how it will be shaped by the actions it takes during training — a behavior the researchers call a precursor to alignment faking, while explicitly stating they do not think the model is alignment faking.
Selected posts from the edition, with links to the originals.
"Day 11: We've landed a series of changes and more compute to very significantly decrease the occurrence of the disruptive 'Selected model is at capacity. Please try a different model'. When it does happen, it should also eventually just continue by itself and not just block on"
View post on X ↗"New voices. Better accents. And less 'checking...' GPT-Live-1.1 brings these updates to Voice in ChatGPT, Codex, and dots, with new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese. Rolling out today to all users. Coming soon to the API."
View post on X ↗"We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon — stay tuned! Lots of details in our technical blog post."
View post on X ↗"We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer. SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM. Available now under Apache 2.0."
View post on X ↗"We've opened an office in Singapore—our fifth in the Asia-Pacific region. We've also signed an MOU with the Government of Singapore on AI safety and responsible adoption."
View post on X ↗"Over the last few years, we have made significant progress on prompt injection for Claude models. As agents do more, it is increasingly urgent that we solve it across the industry. Really excited to partner with Gray Swan to host this competition."
View post on X ↗"We want to support the mathematics community in turning progress in AI into deeper human understanding and new discoveries. To support that goal, we are contributing $5 million to a new Simons Foundation initiative to help mathematicians examine, understand, and build on results"
View post on X ↗"Check out @tonbistudio's guide to using the new Reverse Engineer Anything plugin in Hermes Agent!"
View post on X ↗"Exciting update: There's a new leader in the Agent Arena among open models. MiMo V2.6 Pro by @XiaomiMiMo has landed the #1 spot with +3.8% net improvement!"
View post on X ↗"The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious"
View post on X ↗