Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

ID: 35937
Status: summarized
Published: 15 Oct 2026, 8:00 AM
Fetched: 16 Oct 2026, 2:43 AM
Provider: Hugging Face Blog
Category: developer-ai
Original URL: https://huggingface.co/blog/meta-muse-spark-open
Source URL: https://huggingface.co/blog/feed.xml

Summary

Score: 6.5
Created: 16 Oct 2026, 2:43 AM
Tags: open-weights multimodal mixture-of-experts ai-agents model-release
Audience: developersai_ml_learnersai_agent_users

What happened

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model on Hugging Face with roughly 816B total parameters, ~42B active per token, and a context window up to 1M tokens. It accepts text, images, audio, and video as input but outputs text only, and Meta says it was trained for agentic coding tasks like multi-file refactors, long debugging sessions, tool calling, and browsing with images. Checkpoints ship in BF16 and NVFP4 formats under Muse-Spark-Open-816B-A42B, with 256 experts per MoE layer and 8 selected per token; it shares a tokenizer and chat template with the smaller dense Muse Glimmer released in August.

Why it matters

The deployment-relevant detail is the NVFP4 checkpoint alongside BF16: at 42B active parameters per token and a 1M-token context, the quantized checkpoint is likely what decides whether self-hosting is feasible versus staying on an API. If your agent loop benefits from screenshot, voice-note, or screen-recording input, this is one model that handles all three in a single interface, but output is text-only, so any image or audio generation still needs separate tooling. Note that 'most capable US open-weights model to date' is Meta's own claim in this post and is not benchmarked here, and the post had only 9 upvotes at capture, so independent evaluation is still thin.

Discussion angle

For agentic coding, does 42B active params with NVFP4 actually beat calling a hosted model on cost and latency at 1M context, or is the practical win just the multimodal input (screenshots and screen recordings) inside the agent loop?