Meta releases open-weight Muse Spark Open: 816B MoE, 42B active, 1M context
QGallouedec · x · 2026-10-16
Meta released its open-weight flagship Muse Spark Open, with a Hugging Face guide on running and fine-tuning it:
- Architecture: MoE with 816B total params, 42B active per token; 256 experts per layer, 8 selected per token; context window up to 1M tokens
- Modalities: accepts text, image, audio, and video input (reads screenshots, listens to voice notes, watches screen recordings), outputs text; the most capable US open-weights model to date, following Muse Glimmer
- Positioning: trained for agentic coding — codebase reading, multi-file refactors, long debugging sessions, tool calling, plan-act-observe self-correction, and web browsing with images
- HF post includes setup instructions, fine-tuning how-tos, and example code
Related event: Meta Open-Sources Muse Spark Open, an 816B-Parameter MoE Model(44 posts)→
More from coding & agent
- Manus partners with OPPO for Find X10 agent mode; Perplexity open-sources SPACE sandbox — heyneighbor · 2026-10-16
- Lindy learns tasks by watching once, adds phone-call features and reshuffles pricing — heyneighbor · 2026-10-16
- Genspark Claw outage blamed on Microsoft, adds iMessage/WhatsApp/Slack/LINE access — heyneighbor · 2026-10-16
- Perplexity Computer Ships GPT-6 Astra for Max, Open-Sources SPACE Sandbox Under Apache 2.0 — heyneighbor · 2026-10-16
- Indie agents weekly roundup: open-source sandbox, smartphone partnership, Mac launch, two fundings — heyneighbor · 2026-10-16
- NanoClaw ships first calendar-versioned stable release; Hermes Agent v0.22 adds one-click local models — heyneighbor · 2026-10-16