Qwen3.8 27B benchmarks on NVIDIA RTX Spark: 19.3 tok/s decode, 100% accuracy
draginol · x · 2026-10-19
Blogger draginol shared local inference benchmarks for Qwen3.8 27B Q4KM (35K context, Locama GGUF) running on an NVIDIA RTX Spark N1X (Surface Laptop Ultra with 6144-core Blackwell RTX GPU, 20-core CPU, 64GB RAM, Windows):
- Prefill: 664 tok/s at 8K context
- Decode: 19.3 tok/s while answering
- First token: 1.98s with a 1K prompt
- Accuracy: 100% (80 of 80 items), with some thermal throttling observed during the run
He calls Qwen3.8 27B the gold standard in this class, having previously tested a 12B model on the same machine. A useful public reference point for running 27B-class models on consumer Blackwell hardware.
More from Infra
- IBM exec: US AI leadership hinges on permitted power and domestic robotics scale vs China — ingliguori · 2026-10-19
- Matthew Green: What if all frontier AI models go down in a major incident? — matthew_d_green · 2026-10-19
- HotKV semantic cache cuts Claude tokens 70%, but answer quality dips until threshold 0.88 — HotKV-Ltd · 2026-10-19
- Critics question Anthropic compute shortage amid virtual 'scientists' claims — gerardsans · 2026-10-19
- David Sacks: payback period is the North Star metric for AI infra — CoreWeave recaps in 2.5 years on 5-year deals — DavidSacks · 2026-10-19
- Gemma 4 12b on NVIDIA RTX Spark laptop: 31.6 tok/s decode, 100% accuracy — draginol · 2026-10-19