Meta's Muse Spark Open hits ~150M tokens/s per MW on NVIDIA GB300 in early benchmarks
ArtificialAnlys · x · 2026-10-24
Artificial Analysis published first results for Meta's newly released 816B (42B active) open-weights model Muse Spark Open on its AA-AgentPerf leaderboard.
- Scores 39 on the AA Intelligence Index; MoE with 62 layers, 46 using 2,048-token sliding-window attention, cutting KV cache needs at 1M context by 75% vs full attention
- In throughput-oriented configs, NVIDIA B300 and GB300 both reach 150M tokens/s per MW; GB300 leads at higher output speeds (108M vs 92M at 100 tok/s), with the gap expected to widen in multi-node disaggregated serving
- Tuned SGLang configs deliver 1.6-1.8x the throughput per MW of launch-week cookbook configs on the same B300 node, at the cost of longer time to first token
- AMD MI355X hits 70M tokens/s per MW (0.9x B200); AMD's Oct 22 recipe update lifted per-MW performance 1.7x
Weekly leaderboard updates and deeper reviews of kernel/inference-engine optimization are planned.
Related event: Meta's Muse Spark Open Hits 150M tokens/s per MW in Benchmarks(2 posts)→
More from Infra
- humans& ships 700B-param collaborator model Ampersand, social product Boba, and open-sources FTL decode — ericzelikman · 2026-10-24
- Stripe reportedly buying OpenRouter for ~$7.5B as AI usage billing becomes the prize — FinanceYF5 · 2026-10-24
- Redditor pushes back on local AI FUD: real inference speeds are 5x what the video shows — wojtek15 · 2026-10-24
- NVIDIA ships BioNeMo Inference Runtime v0.2 with RFD4-Proteina support — AllThingsApx · 2026-10-24
- ISTA-DASLab ships GSQ-RCO quantized DeepSeek-V4.1-Flash GGUFs, llama.cpp support still pending — Gold_Coconut9777 · 2026-10-24
- Sauers_ Mocks OpenAI Over Unstudied Multi-Agent Hardware — Sauers_ · 2026-10-24