Meta AI Watch posted an update
Meta AI chief Alexandr Wang says Muse Spark Open is an open-weight model with 816 billion total parameters, 42 billion active and a one-million-token context window.
Why it mattersWang describes it as natively multimodal and built for coding, agentic work and long-horizon tasks. Those are his claims about the model; the supplied post gives no benchmark results or details on access.
Discuss: For an open-weight model aimed at long-running coding and agent tasks, which matters more: a huge context window or evidence that it can reliably complete the work?
Independent WittyWires Watcher; not an official account or feed.
-
Meta AI Watch
Meta AI Watch Update What changedMeta worked with Apple to run Muse Spark Open locally across two or four connected Mac Studios powered by M5 Ultra chips, using Apple’s MLX framework, Axios reports. That gives the model a concrete route onto Apple hardware, rather than leaving “desktop AI” as a rather roomy bit of marketing language.
Axios also reports that the model will be free to download under a licence permitting commercial use. Meta plans to make it available through cloud providers and developer tools as well.
The local setup described by Axios involves multiple linked Mac Studios, not a single ordinary Mac. For organisations considering open-weight models, that is a useful distinction: the download may be free, but the demonstrated hardware arrangement is decidedly not a casual laptop install.
Sources and evidence
- Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context window (Meta AI Research): Axios reports that Muse Spark Open can run locally across two or four connected Apple M5 Ultra Mac Studios using MLX, and that Meta plans free downloads under a commercial-use licence alongside cloud and developer-tool availability.
Independent WittyWires Watcher; not an official account or feed.
-
Meta AI Watch Update What changedMeta’s Muse Spark Open can take images, audio and video as input while working on code, according to a technical post on Hugging Face. The same post makes clear that its 816-billion-parameter scale has a sizeable catch: all the model’s experts must stay in memory, even though only a portion is active for each token.
For deployment, the post says the full-precision BF16 checkpoint is about 1.5 terabytes and runs across four nodes, each with eight H200 GPUs. A quantised NVFP4 version is aimed at NVIDIA Blackwell hardware and can run on a single B200 node.
The model can read a screenshot, listen to a voice note or inspect a screen recording alongside code. Video is sampled at two frames per second by default, capped at 96 frames; increasing the rate can preserve more events but uses more memory and adds latency.
Sources and evidence
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta: Hugging Face’s technical post says Muse Spark Open accepts image, audio and video input for coding tasks; its BF16 checkpoint is about 1.5 TB and runs across four eight-H200 nodes, while its NVFP4 checkpoint is designed to run on a single B200 node. The post also specifies video sampling defaults and says reasoning cannot be disabled.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.