NVFP4 DeepSeek on 4x DGX Spark reproduced: 2.3x vLLM at one user, 1.3x under load
AIFlow_ML · x · 2026-11-03
Local AI Daily's Nov 3 roundup features an independent reproduction of NVIDIA's NVFP4 DeepSeek-V4.1-Flash on 4× DGX Spark, comparing TensorFold and vLLM TP4 recipes on the same hardware and checkpoint (measured, raw outputs public).
- Decode (code): 71.0 vs 33.6 tok/s
- Total throughput: 62.1 vs 27.6 tok/s at 1 user; 144.6 vs 114.8 tok/s at 8 users
- First token on 8K prompt: 2.2 s for both
- The speedup shrinks under load: 2.3× at 1 user, only 1.3× at 8
Also covers Mac vs Spark under load; items are labeled MEASURED / VENDOR-CLAIM / ESTIMATE with licence annotations.
More from Infra
- Apple Q4: $114.5B revenue up 12%, AI/Siri spending to squeeze margins — mitchdeg · 2026-11-03
- Darkbloom passes 1T tokens as Mac compute network pivots to CI/CD — gajesh · 2026-11-03
- OpenRouter Hits 100T Weekly Tokens; Surplus Jumps to 2.87T/Week in 4 Weeks — 0xJeff · 2026-11-03
- Claude Code caches are model-specific: switching models, compacting and idle time all bust the cache — shoumikgoswami · 2026-11-03
- Strata update doubles local inference speed for quad-R9700 setups: 70 to 140-160 tok/s — poofph · 2026-11-03
- Space-based distributed training gains traction as researchers push orbital datacenter vision — MillionInt · 2026-11-03