Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal MoE model
In short: Meta published Muse Spark Open, an open-weight mixture-of-experts model with about 816 billion total parameters and 42 billion active per token. It accepts text, images, audio and video as input and generates text, with a context window of up to 1 million tokens. Meta positions it as a coding-optimized, agentic model built on top of the smaller Muse Glimmer architecture, calling it the most capable US open-weights model released so far.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Mixture-of-experts architecture: 256 experts per layer, 8 selected per token, ~816B total / ~42B active parameters
- 2Adds audio and video input on top of Muse Glimmer's text/image support, via a dedicated audio encoder (12.5 tokens/second)
- 3Context window extended to up to 1,000,000 tokens, including image/video/audio tokens
- 4Released in BF16 (~1.5TB) and NVFP4 (4-bit, for NVIDIA Blackwell GPUs) checkpoints
- 5Built-in tool calling and configurable reasoning effort (low/medium/high/xhigh), with reasoning that cannot be fully disabled
- 6Trained for agentic coding workflows: multi-file refactors, long debugging sessions, and web browsing with images
| Parameter | Before | Now |
|---|---|---|
| Architecture | Muse Glimmer: dense feed-forward network per layer | Muse Spark Open: Mixture-of-Experts, 256 experts per layer, 8 selected per token |
| Model size | Not specified | ~816B total parameters, ~42B active per token |
| Context window | Not specified | Up to 1,000,000 tokens |
| Inputs | Not specified (Muse Glimmer: no audio encoder) | Text, images, audio, and video |
| Output | Not specified | Text |
| Checkpoint formats | Not specified | BF16 and NVFP4 (4-bit, for NVIDIA Blackwell GPUs) |
| Tool calling | Not specified | Supported, text-only and multimodal |
Why it matters
This is a significant open-weight release that pairs frontier-scale reasoning and long context with native audio and video understanding, something most open models still lack. For teams building agents, it offers a single model that can read code, look at screenshots, listen to audio and call tools, reducing the need to stitch together separate STT and vision models.
What it means for AI agents and contact centers
The native audio encoder (accepting raw 24kHz mono waveforms) is the most relevant piece for voice AI and contact-center work: it is worth comparing how well it transcribes or reasons over call audio versus a dedicated STT+LLM pipeline, checking latency and cost at the required reasoning-effort level, and testing whether it handles smaller languages reliably, since none of this is specified in the source for languages like Lithuanian.
🧪 Worth Testing
Direct audio input combined with tool calling and long context could simplify agentic voice workflows, but real-world latency, hosting cost (816B params, multi-GPU requirements) and non-English language quality need hands-on verification before considering it for production call-handling use cases.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta“15 Oct 2026, 03:00
- AWS Machine LearningOfficialOriginal article →„Meta’s Muse Spark Open now available in Amazon SageMaker JumpStart“15 Oct 2026, 20:43
- NVIDIA DeveloperOfficialOriginal article →„Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72“15 Oct 2026, 21:00
- Published by source
- 15 Oct 2026, 03:00
- Found by our system
- 15 Oct 2026, 20:41
- Summary generated
- 15 Oct 2026, 20:43
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy