Meta open-sources Muse Spark: 816B params, 1M context, built for agents

Meta has released the weights for Muse Spark Open, which it calls the most capable US open-weights model so far. It is a mixture-of-experts model with 816B total parameters, 42B active per token (8 of 256 experts per layer) and a 1M-token context window. It reads text, images, audio and video, and outputs text. It shares a base model with Muse Spark 1.3 and adds extra safety post-training so it can run without Meta's system guardrails.

It is tuned for agentic coding: multi-file refactors, debugging, tool calling and long-horizon tasks. On Meta's own numbers it posts 89.2 on MCP-Atlas and 76.4 on Terminal-Bench 2.1, but it trails Kimi K3 and GLM 5.3 on Terminal-Bench and on the Artificial Analysis Intelligence Index (39 vs 44–45). Nobody has verified the scores independently yet.

The weights come in BF16 and NVFP4 on Hugging Face and run on vLLM, SGLang, llama.cpp and MLX. The NVFP4 version still needs about 700GB of memory. The Muse Model License is not fully open: products must say "Built with Muse Spark Open," and AI-as-a-service providers with more than $20M in revenue must sign a separate deal with Meta.

Why it matters: Meta is back at the frontier of open weights, though Chinese labs still lead on the coding benchmarks.