RELEASE · MODELS · #2061
Meta releases Muse Spark Open — 816B-parameter open-weight multimodal MoE model
Meta published Muse Spark Open, an open-weight mixture-of-experts (MoE) model with ~816 billion parameters (≈42B active per token) that accepts text, images, audio, and video and supports up to a 1,048,576-token context window. Meta says the model is optimized for agentic coding workflows (reading, writing, debugging code and tool calls); checkpoints are available in BF16 (safetensors, ~1.5 TB) and an NVFP4 4-bit format tuned for NVIDIA Blackwell GPUs.
KEY POINTS
- Meta published Muse Spark Open, an open-weight mixture-of-experts (MoE) model with ~816 billion parameters (≈42B active per token) that accepts text, images, audio, and video and supports up to a 1,048,576-token context window.
- Meta says the model is optimized for agentic coding workflows (reading, writing, debugging code and tool calls); checkpoints are available in BF16 (safetensors, ~1.5 TB) and an NVFP4 4-bit format tuned for NVIDIA Blackwell GPUs.
- An open-weight 816B multimodal MoE with 1M-token context and practical low-bit checkpoints materially broadens access to frontier multimodal and agentic coding capabilities and affects deployment trade-offs and ecosystem tooling.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Meta released Muse Spark Open, an open-weight mixture-of-experts (MoE) model with ~816 billion parameters and ~42 billion active parameters per token (top-8 routing across 256 experts). The model is multimodal (text, images, audio, video), supports up to a 1,048,576-token context window, and is presented as optimized for agentic coding workflows (reading/writing/debugging code and tool calls). Checkpoints include an NVFP4 4-bit format (published to Hugging Face) optimized for NVIDIA Blackwell (GB300 NVL72) hardware; the model was trained on NVIDIA GB300 NVL72 and uses a 1.8B perception encoder for image/audio/video inputs.
WHY NOW
An open-weight 816B MoE with 1M‑token context, multimodal inputs, and a practical NVFP4 checkpoint optimized for NVIDIA Blackwell hardware changes access and deployment trade-offs: it enables experimentation with long-context, multimodal, agentic coding workflows at frontier scale using downloadable weights and Blackwell‑optimized low‑bit formats.
WHO IS AFFECTED
Directly affected parties identifiable in the sources: developers and teams doing agentic coding workflows and long, multi-file code tasks; researchers and engineers experimenting with multimodal/long‑context models; operators and vendors using or targeting NVIDIA GB300 NVL72 (Blackwell Ultra) GPUs; and Hugging Face users downloading and running the published checkpoints.
CONFIRMED
Explicit source-backed facts: - Model name: Muse Spark Open (Meta) [2442,2439]. - Total parameters ≈816 billion; active parameters ≈42 billion per token via top-8 routing over 256 experts [2442,2439]. - Model modalities: text, images, audio, and video; includes a 1.8B perception encoder for multimodal inputs [2442]. - Context window: up to 1,048,576 tokens [2442,2439]. - Architecture details in source: 62 layers, hidden dimension 8192, experts are FFNs projecting 8192→4096→8192; experts hold ~97% of total parameters [2439]. - Checkpoints: repositories named Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4 on Hugging Face; NVFP4 weights downloadable from Hugging Face [2439,2442]. - NVFP4 checkpoints are presented as optimized for NVIDIA Blackwell Ultra (GB300 NVL72) and NVIDIA provides a downloadable NIM for deployments [2442]. - NVIDIA reports training on GB300 NVL72 and quotes throughput claims (up to 50,000 tokens/sec per GPU and up to 600 tokens/sec per user for agentic workloads on GB300 NVL72) [2442].
UNCERTAIN
Missing or conflicting evidence (explicitly labeled): - License and redistribution terms for the published checkpoints are not stated in the supplied excerpts (missing evidence). - Presence, format, and size of any BF16 safetensors checkpoint (and exact storage footprint) are not documented in the supplied excerpts (missing evidence). - Independent third‑party benchmark verification of NVIDIA/Meta throughput, accuracy, or tool‑calling reliability claims is not provided; NVIDIA’s performance and accuracy claims are manufacturer‑reported (single-source claim) [2442] (limited/one-sided evidence). - Operational performance and memory requirements on non‑Blackwell GPUs (e.g., older architectures) are not described (missing evidence). - Training data composition, safety mitigations, and model limitations are not described in the supplied excerpts (missing evidence). - Any differences between the NVFP4 checkpoint behavior and full‑precision behavior (e.g., fidelity or degradation) are not quantified in the excerpts (missing evidence).
WHAT TO WATCH
Concrete observable signals to monitor next (sources cited): - Hugging Face repository activity for Muse-Spark-Open-816B-A42B and -NVFP4: downloads, releases, README/model card updates, license file and file sizes [2439]. - Third‑party benchmarks and independent evaluations comparing throughput, accuracy, and tool‑calling reliability to published claims (search for benchmark reports citing Muse Spark Open) [2442]. - NVIDIA NIM package releases, documentation, and deployment guides for Muse Spark Open on GB300 NVL72 and any benchmarks/recipes provided by NVIDIA [2442]. - Community issues, forks, and fine‑tuning recipes posted on Hugging Face or GitHub showing practical memory/latency on non‑Blackwell hardware [2439]. - Official model card, security/safety documentation, or Meta publications clarifying training data, license, and known limitations (model card updates on Hugging Face or Meta posts) [2439].
WHY IT MATTERS
An open-weight 816B multimodal MoE with 1M-token context and practical low-bit checkpoints materially broadens access to frontier multimodal and agentic coding capabilities and affects deployment trade-offs and ecosystem tooling.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Muse Spark Open is an open-weight mixture-of-experts (MoE) model with approximately 816 billion parameters.
SUPPORTEDChecked 2026-10-15 · 2 supporting
The model activates about 42 billion parameters per token via top-8 routing across 256 experts (router selects 8 experts; 256 experts per MoE block).
SUPPORTEDChecked 2026-10-15 · 2 supporting
Hugging Face: each MoE block has 256 experts and the router selects 8 experts per token; HF notes ≈42B active per token.
SUPPORTS NVIDIA DeveloperMEDIA · 2026-10-15NVIDIA: Muse Spark Open activates 42B of 816B total parameters per token through top‑8 routing across 256 experts.
Muse Spark Open accepts text, images, audio, and video as inputs and processes interleaved modalities natively.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Hugging Face: the model brings text, images, audio, and video into a single model and generates text about different kinds of input.
SUPPORTS NVIDIA DeveloperMEDIA · 2026-10-15NVIDIA: Muse Spark Open adds audio and video perception alongside image and document understanding, processing interleaved modalities natively.
The model uses a perception encoder (reported as about 1.8B parameters by NVIDIA) for multimodal inputs.
SUPPORTEDChecked 2026-10-15 · 1 supporting
SOURCES & TIMELINE
4Mixture-of-experts (MoE) model Muse Spark Open activates 42B of 816B total parameters per token through top-8 routing across 256 experts. Trained on NVIDIA GB300 NVL72 Blackwell Ultra for agentic coding workflows, it delivers higher first-attempt accuracy, reliable tool calling, and deep codebase understanding across repositories. A 1M token context window holds an entire project including complex dependencies, leg…
Muse Spark Open is an open-weight model from Meta. The mixture of experts model brings text, images, audio, and video into a model with approximately 816 billion parameters, around 42 billion active per token, and a context window of up to one million tokens. It generates text, giving developers a common interface for asking questions about different kinds of input. Following Muse Glimmer, Muse Spark Open introduces…
Meta has launched what it calls the “most capable” open-weight model in the United States. Dubbed Muse Spark Open, the model also has the distinction of having been designed to work with Macs. The new model, which leverages 816 billion parameters and a 1 million token context window, claims to offer superior performance to other popular open weight models, according to evaluation results released by Meta on Thursday…
After its Llama series, Meta is returning to open-weight territory with Muse Spark Open. Meta calls it the most capable US open-weight model, built for coding and agentic tasks, with 816 billion parameters (42 billion active), up to one million tokens of context, and support for text, image, video, and audio. On benchmarks like JobBench, MCP-Atlas, HLE-Diamond, and ZeroBench, Meta says it scores well ahead of Qwen 3.…