DGX Spark benchmarks: DeepSeek V4 Flash passes 900K-token prompt locally
jtsaint333 · reddit · 2026-08-29
The author tested several new models on four DGX Sparks (two ConnectX-7 pairs) with first-hand local results:
- DeepSeek V4 Flash (2-way TP2, vLLM, NVFP4 MLA KV, MTP5) served 1M context and passed a 899,994-token prompt — but was slow: high thinking took over an hour for the 24-task batch with 20/24 API success.
- Qwen3.8 Flash Next scored best: 22/24 with medium thinking; thinking-off finished the whole batch in 1:22 with only 8,861 output tokens.
- Qwen3.8-27B delivered the highest throughput (190 delivered tok/s, thinking off); low thinking cost 3.82x wall time for one extra pass.
- Qwen3.6-35B-A3B on a single card hit 784 tok/s aggregate at c16, scoring 11/12 on a programming subset with 32K context.
Operational takeaways outweigh one-run scores: high/xhigh thinking was counterproductive here — xhigh used 3x the tokens of medium and scored worse. The author stresses these are deployment operating points, not an architecture-normalized leaderboard.
More from Infra
- Next-gen GraalVM benchmarks: 4x faster startup, 2x less memory, no build step needed — RealGeneKim · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark AI laptops — tomwarren · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark laptops — The Verge AI · 2026-10-15
- Global smartphone shipments plunge 22% in worst quarterly decline on record as AI demand fuels memory costs — Polymarket · 2026-10-15
- SK Hynix drops 11% again, now below 3x FY2027 earnings as AI memory selloff deepens — JOBhakdi · 2026-10-15
- ASML Sells Zero Lithography Machines into Europe as US Hosts 75%+ of AI Compute — NinaDSchick · 2026-10-15