Meta's Muse Spark Open hits 39 on AA index, ~150M tokens/s per MW on GB300/B300
ArtificialAnlys · x · 2026-10-24
Artificial Analysis published first AA-AgentPerf results for Meta's newly released 816B (42B active) open-weights model Muse Spark Open:
- Up to 150M tokens/s throughput per MW of power on 4x NVIDIA GB300 or B300 GPUs.
- Scores 39 on the Artificial Analysis Intelligence Index.
- MoE with 62 layers, 46 of which use 2,048-token sliding-window attention.
AA also launched AA-AgentPerf, a hardware benchmark for the agent era that replays real coding agent trajectories, measuring long context, tool-calling delays, and KV cache reuse—trajectories grow past 100K tokens with only new tokens prefilled. The benchmark is open for submissions, with serving configs from DeepSeek V4.1 Flash, Kimi K3, GLM-5.2/5.3 and others already listed.
Related event: Meta's Muse Spark Open Hits 150M tokens/s per MW in Benchmarks(2 posts)→
More from Infra
- NSA says take PHP systems offline if unpatchable; some Cisco fixes not due until Feb 2027 — shashib · 2026-10-24
- If It Flies, Floats or Finetunes, Renting Is Cheaper — yaroslavvb · 2026-10-24
- Lightspeed's Amber Yang Calls for Startups to Build the "Standard Oil of Compute" — theamberyang · 2026-10-24
- LTX 2.5 distilled 22B Q4 runs locally on an 8GB M1 MacBook Air — 36s of video per ~32min render — iJohnnyRobert · 2026-10-24
- Directionally Local launches private coding-agent infra that runs on edge hardware you own — sahilypatel · 2026-10-24
- What $70-100K of hardware buys for local GLM5.3 and DeepSeek with 10-20 concurrent users — whoami-233 · 2026-10-24