Open-weights Muse Spark on B300 costs $0.12 vs $0.88 per correct task, 7.1x cheaper
ryanshrout · x · 2026-10-20
Ryan Shrout shared Signal65 PINNACLE results for Muse Spark Open: on an 8× NVIDIA B300 node, the NVFP4 build completes a correct enterprise task for $0.12 vs $0.88 on Meta's Muse Spark 1.3 API — 7.1x cheaper. With DSpark speculative decoding, NVFP4 holds 112 concurrent agents vs 24 for BF16 (4.7x), losing under 3% of the PINNACLE Model Score with identical 97.2% retrieval fidelity on a 128K corpus. The hosted model still makes 55% fewer weighted errors, almost all on the hardest agentic tier. PINNACLE is an agentic benchmark scoring "correct work" with procedurally regenerated sandboxes, code-verified scoring, and live agent sessions; NVIDIA and AMD executives have endorsed it.
More from Infra
- NVIDIA's Ian Buck explains how AI factories can flex power use to relieve the grid — nvidia · 2026-10-20
- p99 container startup latency cut 63%, saving ~19 hours of compute daily on Cloud Run — DanielLockyer · 2026-10-20
- Late Interaction in 2026: Experiment Shows Single-Vector Search Ranks Wrong Answers Higher — IgorCarron · 2026-10-20
- AMD, NVIDIA, and Google All Ship Major vLLM Inference Optimizations in Three Days — PyTorch · 2026-10-20
- XFreeze Claims SpaceX Orbital Data Centers Are the Endgame for Scaling Intelligence — XFreeze · 2026-10-20
- Intel's Crescent Island GPU D270: A 160GB PCIe 5.0 Inference Card — lemire · 2026-10-20