Meta Releases Muse Spark Open: 816B/42B Model With 1M Context, Tops Agentic Benchmarks
TorturedPoet30 · reddit · 2026-10-16
Meta released Muse Spark Open, billed as the most capable US open-weight model to date, built for coding and agentic work:
- Specs: 816B total / 42B active parameters, 1M token context, natively multimodal across text, image, video, and audio
- Positioning: long-horizon tasks, complex software engineering, multi-step reasoning
- Benchmarks (vs Inklings, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8, GLM 5.3, Kimi K3): leads on JobBench 61.8, MCP-Atlas 89.2, HLE-Diamond 26.1, SciCode 58.5, AA-Omniscience 25; Terminal-Bench 2.1 76.4 and DeepSWE 1.1 60 trail GLM 5.3 and Kimi K3; multimodal ZeroBench 50.0 and BabyVision 80.2 lead the field
- Weights already hosted by Unsloth on Hugging Face (Muse-Spark-Open-816B-A42B)
Strong on agentic and reasoning evals; pure coding benchmarks still behind GLM 5.3 and Kimi K3.
Related event: Meta Open-Sources Muse Spark Open: 816B-Parameter MoE with 1M Context(54 posts)→
More from Models
- Gemini 4 Argon rolls out globally in Google Search's AI Mode for Pro/Ultra subscribers — OnlineInference · 2026-10-16
- Dev Builds Haunted Escape Room to Test Reflection AI's Beam, Solves It in 30 Seconds — wandb · 2026-10-16
- Claude Opus 5.5 fast mode exits preview: 2.5x faster at $8/$40 per million tokens — ClaudeDevs · 2026-10-16
- NVIDIA ships Nemotron 3.5 Bolt: 74B-A9B model in 24GB of memory — NVIDIAAI · 2026-10-16
- User's Opus 5.5 High Blender 3D experiment crashes their Mac with hilarious results — imjustnewatai · 2026-10-16
- Apollo Research: GLM-5.3 Spontaneously Reasons About How Its Actions Shape Future Training — jd_pressman · 2026-10-16