According to an NVIDIA technical blog post dated Oct 15, Meta launched Muse Spark Open, a coding-optimized 816B-parameter mixture-of-experts model with 42B active parameters and a 1M-token context window. The model routes each token to eight of 256 experts, and Meta produced NVFP4 checkpoints aimed at Blackwell Ultra. NVIDIA says a 1.8B-parameter perception encoder processes interleaved text, images, video frames and audio in a single pass, and that the model can search the web, spawn and coordinate subagents, and use computer control across desktop, mobile and browser. On a GB300 NVL72 rack (72 Blackwell Ultra GPUs, 130 TB/s NVLink), NVIDIA reports up to 50K tokens per second per GPU and up to 600 tokens per second per user on agentic workloads with speculative decoding. NVIDIA lists SGLang, vLLM and TensorRT LLM as supported serving stacks, with post-training through NeMo AutoModel and Megatron Bridge.
Our takeThe interesting claim is not the parameter count but the ratio: 42B active out of 816B keeps serving cost near a mid-size dense model while the window holds a whole repository. Two caveats: the throughput figures come from NVIDIA's own four-GPU configuration with speculative decoding, not an independent harness, and the post asserts higher first-attempt accuracy without numbers or a named baseline. Long-context coding agents tend to fail on retrieval precision, not window size.