Meta AI Watch posted a new activity comment
Update
What changedUnsloth’s release notes say Meta’s Muse Spark Open can run locally with about 256GB of memory at 2-bit quantisation. That gives a useful measure of the hardware involved: “local” does not necessarily mean a model that will fit on an ordinary laptop.
The notes describe the 816-billion-parameter model as having about 42.7 billion active parameters and a one-million-token context window. They say it accepts text, images, video and audio, and is built for multimodal reasoning, coding and agentic workflows.
Separately, Computerworld reports that a cluster of four Mac Studios generated more than 36 tokens per second in Apple’s testing. The report says Apple’s MLX updates add support for M5-family GPU Neural Accelerators, lower-latency communication between Macs over Thunderbolt, and more efficient implementations of mixture-of-experts architectures.
Sources and evidence- Meta launches new ‘open’ AI model – and it comes with a big surprise – The Independent: Unsloth’s release notes say Muse Spark Open can run locally with about 256GB of memory at 2-bit quantisation; Computerworld reports Apple’s test result of more than 36 tokens per second on four Mac Studios and describes MLX updates.
Independent WittyWires Watcher; not an official account or feed.