News
Meta released Muse Spark Open, an 816B open-weight multimodal model with a 1M-token context window, and NVIDIA detailed GB300 NVL72 deployment.
Meta’s release of Muse Spark Open and NVIDIA’s deployment details mark a push to make large, multimodal models more practical for long-horizon coding and agentic workflows. The companies say the model is designed to keep an entire software project in context while running on NVIDIA’s GB300 NVL72 system for higher-throughput inference.
Evidence and context
According to Meta’s announcement on October 15, 2026, Muse Spark Open is an open-weight, coding-optimized mixture-of-experts model with 816 billion total parameters and about 42 billion active parameters per token. Meta says the model uses a 1,048,576-token context window and is aimed at agentic coding, multistep reasoning and complex software engineering. NVIDIA’s October 15, 2026 technical blog says the model was trained on GB300 NVL72 Blackwell Ultra infrastructure and is tuned for agentic coding workflows.
The context window is central to the pitch: Meta says it can hold an entire project, including dependencies, legacy code and thousands of files, without breaking a task into pieces. NVIDIA adds that the sparse MoE design routes each token to eight of 256 experts, so interconnect bandwidth matters for realized throughput. NVIDIA says the GB300 NVL72 combines 72 Blackwell Ultra GPUs into a single rack-scale NVLink domain with 130 TB/s of communication bandwidth.
Meta also says Muse Spark Open is natively multimodal, handling text, images, audio and video in one pass. The company says the model’s 1.8B perception encoder processes interleaved inputs without separate modality pipelines. NVIDIA’s blog frames that as useful for workflows that mix screenshots, diagrams, error logs and screen recordings during debugging or code review.
NVIDIA says the model uses NVFP4 checkpoints optimized for Blackwell Ultra, reducing memory footprint and supporting larger KV cache for longer concurrent sessions. Meta’s documentation, cited in the supplied background, says the BF16 checkpoint is about 1.65 TB and the NVFP4 checkpoint about 537 GB, before inference overhead and KV cache. Those figures underscore that sparse activation reduces per-token computation, but does not remove the need to load very large model weights.
On performance, NVIDIA says Muse Spark Open running agentic workloads on GB300 NVL72 can deliver up to 50,000 tokens per second per GPU and up to 600 tokens per second per user. The company also says deployment options include SGLang, vLLM and TensorRT LLM, with NVIDIA Dynamo recipes that separate prefill from decode and use KV cache-aware routing. Meta’s licensing terms, as noted in the supplied background, are limited and should not be treated as unrestricted open-source licensing.
A separate item in the provided background says Meta introduced Muse Spark 1.3 on September 2, 2026 with improved agentic and coding performance. The October 15 releases build on that work, but the supplied material does not establish independent comparisons versus rival models beyond vendor claims.
Be the first to comment