Qwen3.8-Flash-Next · ExLlama · 1× DGX Spark
English | 简体中文 | 日本語 · ← All recipes
Source level: ✅ Measured · Date: — · Numbers come from the source. This repo did not measure them again.
Hardware
- 1× DGX Spark (DGX Spark)
- Reference cost: ≈ $6,950
Model
- Qwen3.8-Flash-Next ·
Cruz exllamav3 fork · EXL3 3.05bpw · draft depth 5 · 8-bit KV - Download: Not published. See the source.
Engine
- ExLlama — install: https://github.com/turboderp-org/exllamav3
Launch command
Command not published. See the source.
Key parameters
- Context: 262K (243K needle test 15/15)
- Concurrency: native engine only 1 stream (separate vLLM concurrent lane also exists)
- Tags:
3bit,MoE,long-context,speculative
Expected speed
- code 102.6; prose 71.7; JSON 66.2 tok/s (✅ Measured)
Caveats
- Speed changes with drivers, engine version, context length, and prompt length.
Source & attribution
- https://x.com/yume_arasaki/status/2101741448811229219
- Credit to the original author. This page only summarizes and links. Data under CC BY 4.0.
Source file: recipes/text-022.md