On 2026-10-15 Meta released weights for Muse Spark Open, a natively multimodal MoE (~816B total / ~42B active per token, 1M context) aimed at long-horizon coding and agents. Weights are on Hugging Face and peers; hosted APIs include Together, Fireworks, Groq and OpenRouter. Reported figures include DeepSWE 60%, MCP-Atlas 89.2% and JobBench 61.8% (see Meta’s evaluation methodology).
Key Takeaways
- ✓Scale: ~816B total / ~42B active per token; 256 experts/layer with top-8 routing; 62 layers, hidden 8192; context up to 1,048,576 (HF blog)
- ✓Coding/agents (Meta scorecard / AA, xhigh): DeepSWE 60.0%; SWE-Atlas Codebase QnA 58.0%; MCP-Atlas 89.2%; JobBench 61.8%; Toolathlon-Verified 72.8% (eval methodology)
- ✓Throughput (NVIDIA tech blog, agentic on GB300 NVL72): up to ~50K tok/s/GPU and ~600 tok/s/user; NVFP4 checkpoints published (NVIDIA blog)
- ✓Deploy: Muse Model License with limited commercial terms; SGLang / vLLM / TensorRT-LLM / MLX / llama.cpp; Meta demos two Apple M5 Ultra Mac Studios for on-desk local run
- ✓Third-party index: Artificial Analysis Intelligence Index ~38.9–39 (xhigh, as of 2026-10-07)—behind closed frontier, but a key open-weight reference (AA)
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
Background
Open-weight models are strong at one-shot completion, but long-horizon coding agents still struggle with repo-scale context, multi-step tools, and multimodal inputs. On 2026-10-15 Meta released Muse Spark Open, positioning it as the most capable US open-weight model to date for agentic coding you can self-host or run on desk hardware.
Architecture
Per the Hugging Face engineering post: ~62 layers, hidden 8192, MoE with 256 experts/layer and top-8 routing (~816B total / ~42B active); a ~1.8B perception encoder for images/video plus an audio encoder (~one token per 80 ms). Context is configured to 1,048,576. The Harmony-style chat template routes reasoning, commentary, tools and user replies, with low–xhigh reasoning effort (thinking cannot be fully disabled).
Benchmarks
Meta’s evaluation methodology uses xhigh (512), temp=1.0. Scorecard / third-party figures include DeepSWE 60.0%, SWE-Atlas Codebase QnA 58.0%, MCP-Atlas 89.2%, JobBench 61.8%, Toolathlon-Verified 72.8%, and an Artificial Analysis Intelligence Index near 39 (as of 2026-10-07). NVIDIA reports up to ~50K tok/s/GPU and ~600 tok/s/user for agentic loads on GB300 NVL72—vendor numbers; re-measure in your stack.
Getting started
Weights: HF collection (BF16 and NVFP4). Hosted APIs: Together, Fireworks, Baseten, Modal, Groq, Nebius, OpenRouter. Self-host via SGLang/vLLM NVFP4 recipes; Apple via MLX/oMLX; llama.cpp GGUF quants available. License is the Muse Model License with limited commercial terms.
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.