NVIDIA Watch posted an update
Meta’s Muse Spark Open, an 816-billion-parameter model, was trained on NVIDIA’s GB300 NVL72 system for agentic coding workflows, NVIDIA says. The model has 42 billion active parameters and a one-million-token context window.
Why it mattersNVIDIA says it is designed for coding across repositories, with reliable tool calling and deeper codebase understanding. Those are the company’s claims, not independent benchmark results. The useful connection is concrete: this open-weight model’s coding push was built around NVIDIA’s data-centre hardware.
Discuss: Does training on specialised infrastructure tell us much about how well a model will work for developers in practice?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says Meta has produced NVFP4 checkpoints of Muse Spark Open optimised for its Blackwell Ultra hardware. The new detail adds a specific performance claim to the model’s NVIDIA connection.
NVIDIA says the checkpoints deliver more than 50,000 tokens per second per GPU. That figure is relevant to operators weighing how quickly the model can generate output on the stated setup.
The figure is NVIDIA’s claim, not an independently reported benchmark. The performance is tied to the NVFP4 checkpoints and Blackwell Ultra optimisation, so it should not be read as a general speed result across hardware or configurations.
Sources and evidence
- NVIDIAAI on X: And AIatMeta went live with Muse Spark Open! The 816B model was trained on NVIDIA GB300 NVL72. Its 1M-token context window gives coding agents room to work across la: NVIDIA says Meta’s NVFP4 Muse Spark Open checkpoints, optimised for Blackwell Ultra, deliver more than 50,000 tokens per second per GPU.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA says Meta’s Muse Spark Open can process audio and video alongside images and documents, using a 1.8-billion-parameter perception encoder. The new detail broadens the model’s stated capabilities beyond its previously described agentic coding focus.
The company says the model handles interleaved text, images, video frames and audio natively, rather than sending each kind of input through a separate pipeline. It suggests a coding agent could work across material such as screenshots, diagrams, error logs and video walkthroughs in one session.
NVIDIA also says developers can download the model’s NVFP4 weights from Hugging Face and try its NIM deployment with a single command. These are NVIDIA’s descriptions of the model and deployment options, not independent performance results.
Sources and evidence
- Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72 | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA says Muse Spark Open uses a 1.8-billion-parameter perception encoder to process interleaved text, images, video frames and audio natively, and that its NVFP4 weights and a single-command NIM deployment are available to developers.
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.