On 2026-10-15 Meta released Muse Spark Open open weights: an 816B-total / 42B-active MoE with a Perception Encoder, natively multimodal (text/image/video/audio in → text out), 1M context, aimed at coding and long-horizon agents. Weights are on Hugging Face (BF16/NVFP4), SageMaker, Model Garden and Foundry; APIs include OpenRouter, Together, Fireworks and Groq. Meta calls it the most capable US open-weight model to date and, with Apple, optimized it to run locally across two M5 Ultra Mac Studios via MLX.
Key Takeaways
- ✓Specs: MoE 816B total / 42B active (256 experts, top-8), 62 layers, 1M context; text+image+video+audio in → text out (model card)
- ✓Local: NVFP4 needs ≥~700GB VRAM, BF16 ~2TB; or two Apple M5 Ultra Mac Studios via MLX (Meta blog)
- ✓Agentic benches (xhigh, official table): MCP-Atlas 89.2, JobBench 61.8, OSWorld Verified 80.4, Terminal-Bench 2.1 76.4, DeepSWE 1.1 60, AA Intelligence Index 39
- ✓Serve with SGLang / vLLM / TensorRT-LLM / Transformers / llama.cpp / Unsloth / MLX; Helion kernels claim 1.03–3.95× kernel speedups on NVIDIA/AMD
- ✓License: Muse Model License (display “Built with Muse Spark Open”; MaaS over $20M/yr needs a separate deal); OpenRouter ~$1.25/M in, $4.25/M out
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Meta released Muse Spark Open open weights on 2026-10-15: an 816B-total / 42B-active Mixture-of-Experts with a Perception Encoder, 1M-token context, and native multimodal input (text, image, video, audio → text). Meta positions it for coding and long-horizon agents and calls it the most capable US open-weight model to date.
Architecture (model card): 62 LM layers, hidden 8192, 256 experts with top-8 routing, GQA (128 Q / 16 KV), vocab 202,048; Perception Encoder ~1.8B params with sparse attention factor 4. Inference knobs: reasoning_effort in {low, medium, high, xhigh} (default xhigh); recommended temperature 1.0, top_p 1.0, top_k disabled. The open checkpoint adds safety post-training for standalone hosting.
Official xhigh table on the HF card (not extrapolated): AA Intelligence Index 39; JobBench 61.8, MCP-Atlas 89.2, OSWorld Verified 80.4, Toolathlon Verified 72.8; DeepSWE 1.1 60, SWE-Atlas-QnA 58, Terminal-Bench 2.1 76.4; GPQA Diamond 89.0, HLE-Diamond (no tools) 26.1. Versus Kimi K3 / GLM 5.3 in the same table it trails on AA Index and some coding scores, while leading on MCP-Atlas / JobBench among listed peers.
Get weights from the Hugging Face collection (BF16 and NVFP4). Local NVFP4 needs ~≥700GB VRAM; BF16 ~2TB; alternatively two M5 Ultra Mac Studios via MLX. Hosted APIs include OpenRouter (~$1.25/M in, $4.25/M out), Together, Fireworks, Baseten, Groq, Modal, and Nebius. Muse Model License requires a visible “Built with Muse Spark Open” attribution and a separate Meta agreement for MaaS over $20M aggregate revenue in any 12 months.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.