Model Drop

Muse Spark Open

October 15, 2026 2 min read

Muse Spark Open, which Meta released on October 15, 2026, is the open-weights version of Muse Spark: an 816B-parameter mixture-of-experts model with 42B active, built for coding and agents and small enough to run on two Mac Studios.

Model: Muse Spark Open (meta-models/Muse-Spark-Open-816B-A42B)
Maker: Meta
Ship date: October 15, 2026
Pricing: Free (open weights)
Context: 1M tokens
Weights: Open (Muse Model License, Hugging Face)
Where you can use it:
- Hugging Face weights, BF16 and NVFP4
- OpenRouter, Together AI, Groq, Fireworks
When to reach for it:
- Self-hosted coding and tool-calling agents
- Frontier-class model on two Mac Studios
- Skip it if you resell model access at scale
How it compares:
- Leads GLM-5.3 and Kimi K3 on tool use
- Trails both on coding and overall
What makes it different:
- Text, image, video and audio input
- Extra safety training for open release
Early reactions: Hacker News cheered a big US open-weight release but argued over the license, which makes API sellers above $20M in revenue sign a deal with Meta, and pointed out it scores below GLM-5.3 on Artificial Analysis.
Read more: Introducing Muse Spark Open, Hugging Face model card, Evaluation methodology

The benchmarks

Meta ran Muse Spark Open at xhigh reasoning effort and compared it only with open-weight models (Qwen3.8-2.4T-A95B at xhigh, GLM-5.3 and Kimi K3 at max, plus Mistral Large 4, Nemotron 3 Ultra, Inkling and Beam on some rows); coding scores use each model's preferred harness and Terminal-Bench uses the Terminus 2 agent.

Meta's numbers

MCP Atlas (tool use over MCP): 89.2%. +4.7 over Qwen3.8-2.4T-A95B (84.5%), +5.0 over GLM-5.3 (84.2%), +6.9 over Kimi K3 (82.3%).
JobBench (professional tool use): 61.8. +9.1 over Qwen3.8-2.4T-A95B (52.7), +0.4 over GLM-5.3 (61.4), +7.5 over Kimi K3 (54.3).
GDPval-AA v2 (office task Elo): 1401. −23 under Mistral Large 4 (1424), −212 under Qwen3.8-2.4T-A95B (1613), −252 under GLM-5.3 (1653), −136 under Kimi K3 (1537).
DeepSWE v1.1 (long-horizon agentic coding): 60%. −2 under Mistral Large 4 (62%), +3 over Qwen3.8-2.4T-A95B (57%), −7 under GLM-5.3 (67%), −8 under Kimi K3 (68%).
Terminal-Bench 2.1 (Terminus 2 agent): 76.4%. −5.6 under Qwen3.8-2.4T-A95B (82.0%), −7.5 under GLM-5.3 (83.9%), −8.6 under Kimi K3 (85.0%).
HLE-Diamond (no tools) (hardest expert questions): 26.1%. +8.8 over Qwen3.8-2.4T-A95B (17.3%), +9.7 over GLM-5.3 (16.4%), +3.9 over Kimi K3 (22.2%).
GPQA Diamond (PhD-level science): 89.0%. −4.5 under Qwen3.8-2.4T-A95B (93.5%), −2.7 under GLM-5.3 (91.7%), −4.5 under Kimi K3 (93.5%).
ZeroBench (with tools) (hard visual reasoning): 50.0%. +36.0 over Inkling (14.0%), +9.0 over Kimi K3 (41.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 38.9. −5.9 under GLM-5.3 (44.8), −4.7 under Kimi K3 (43.6), −1.0 under Qwen3.8-2.4T-A95B (39.9), +0.5 over Mistral Large 4 (38.4).