Meta releases Muse Spark Open, its biggest open-weight model yet

Meta's open-weight coding model runs on Mac Studios with Apple's help, while Nvidia pushes new agent models and Anthropic opens its doors to outside safety researchers.

Meta releases Muse Spark Open, its biggest open-weight model yet

Meta released open weights today for Muse Spark Open, an 816-billion-parameter coding and agent model in which only 42 billion parameters activate for any given token. The model tops rivals on tool use but still trails some Chinese open models on coding, and Meta's most capable closed models remain locked behind an API.

Meta says the model can run on two Mac Studios equipped with Apple's M5 Ultra chips, linked together using Apple's MLX framework. Apple CEO John Ternus replied to Mark Zuckerberg's announcement post to congratulate the team, and Meta chief AI officer Alexandr Wang told Axios he was excited to see the two companies cooperating again.

The release is genuinely shipping, with usable weights available on Hugging Face and major clouds.

Nvidia and Mistral both launch new models for agents and forecasting

Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open model built for always-on agents, along with two smaller Bolt models meant for local and edge use. One of the Bolt models uses native 2-bit quantization to fit inside 24 gigabytes of memory, and Nvidia says it is the top open model in its class on PinchBench and second on BrowseComp.

Mistral released Timestral, its first family of time-series forecasting models, aimed at applications like sensor data and financial flows. The largest version has just 209 million parameters, and Mistral says it ranks first on all three major public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, beating the previous best model despite having two-thirds its parameter count. Both companies' models ship today with real deployments and detailed benchmark results behind them.

Nvidia also launched RTX Spark, an Arm-based Windows laptop pairing an Arm CPU with a Blackwell GPU to run AI models locally. Microsoft's Peter Dawoud told The Verge that setup takes two clicks, with Hermes, Perplexity and OpenClaw as the first officially supported agents. The Verge's reviewer found large local models still slower than cloud versions, suggesting they are better suited to tasks left running overnight or in the background.

Anthropic gives outside researchers employee-level access to its safety work

Anthropic gave six researchers from the nonprofit METR laptops, office space and access to internal tools roughly equal to its own risk teams. Anthropic said it cannot block METR's published findings, only require redactions in narrow categories such as customer data or model architecture details, comparing the arrangement to bank regulators embedding examiners inside major banks.

TechCrunch reports the move follows weeks of disclosed incidents of AI agents acting outside their instructions since late August, and fulfills a September proposal from Anthropic CEO Dario Amodei. OpenAI has pledged the same step, and Google, Meta, Nvidia and xAI have all signed a White House accord on independent review.

Google launches Universal Cart and DeepMind hunts for new antibiotics

Google's Universal Cart launched in the United States today across Search and the Gemini app, pulling items from different stores into a single cart, tracking price drops through Google's Shopping Graph of more than 60 billion product listings, and suggesting which card to use for payment. Shoppers can check out through Google Pay or go directly to a merchant's site.

Separately, Google DeepMind used its Antigravity platform and a model called ProtNLM to scan millions of uncharacterized proteins in the UniProt database for candidates with predicted activity against pathogens that cause respiratory infections. Google says the lead candidate is now in wet-lab validation, making it a predicted hit rather than a proven drug.

MLCommons wins funding to build an efficiency benchmark for AI agents

MLCommons was selected to run a $3.5 million, two-year project funded by the National Science Foundation to build MLPerf Agents, a benchmark measuring how efficiently agents complete whole tasks rather than how fast a model answers a single prompt. Initial test areas include software engineering, research and data analysis, and customer service.

Nvidia expands open resources for European researchers

At a summit in Dublin, the European Commission launched a program connecting researchers to compute and training, with Nvidia participating. Nvidia says it has released more than 250 open models, 250 open datasets and over 1,400 open-source projects, and is partnering with national AI computing hubs in France, Spain and Sweden.

White House AI column photo flagged as AI-edited

President Trump posted a before-and-after photo of a White House column comparing an old version to a “restored” one, but Google’s SynthID detector flagged the after photo as made or edited with AI. Reddit users noted the two photos show a different number of steps, while the window reflections are nearly identical between them.

Nvidia's GeForce NOW adds new streaming controls and games

GeForce NOW added new DLSS 4.5 controls, letting members adjust image sharpness, frame generation and ray tracing for games streamed from the cloud without needing high-end hardware. Ten new games joined this week, including Enshrouded's full 1.0 release and Warhammer 40,000: Boltgun 2.

Sources

Hype Brake is hosted by two synthetic AI voices. Every fact is checked against published reporting. How this show is made: https://hypebrake.com/