Meta releases open 816B Muse Spark coding model
Meta has released Muse Spark Open, an open-weight 816B mixture-of-experts model built for agentic coding, adding a 1M-token context and native multimodal perception to the open-model field.
Meta has launched Muse Spark Open, a coding-optimized mixture-of-experts (MoE) model with 816B total parameters, 42B of them active per token. The details come from an NVIDIA technical blog post on deploying the model, not from Meta's own announcement. According to NVIDIA, the model was trained on its GB300 NVL72 Blackwell Ultra systems.
The model routes each token to eight of 256 experts. NVIDIA says it is designed for long-horizon agentic work rather than single-shot completions: multifile refactors, recovering from failing tests, tool calling and coordinating subagents. Its 1M-token context window is meant to hold an entire project in one pass. The post says it uses goal conditioning and context compaction to keep long sessions on track, and that it writes commentary before each action so developers can see what it is doing and why.
A 1.8B-parameter perception encoder handles interleaved text, images, video frames and audio in one pass. NVIDIA also lists web search, including multimodal browsing, and computer use across desktop, mobile and browser.
Meta has published NVFP4 checkpoints tuned for Blackwell Ultra's FP4 Tensor Cores. NVIDIA says the smaller memory footprint leaves room for a larger KV cache, which supports more concurrent long-context sessions. Weights are on Hugging Face. Serving recipes cover SGLang, vLLM and TensorRT-LLM through NVIDIA Dynamo, and a NIM container is available. NeMo AutoModel and Megatron Bridge support fine-tuning.
The post includes no benchmark results, and the license terms are not stated. NVIDIA's throughput figures for the GB300 NVL72 are cut off in the available text.