Worker · RPC
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- Holdsits share of the model
- Enginellama.cpp b1341 · ggml-rpc-server
peak during run 125 °F
Hive Benchmarks · Oct 11, 2026
DeepSeek's own speed helper. On three boxes, it costs 24–32%.
DSpark is DeepSeek's official draft model for V4-Flash-0731: a small model that guesses 5 words ahead from the big model's own state, and the big model checks the guesses. On one machine it is reported to nearly double speed. Split across three boxes, every guess-and-check round has to cross the links between them, and only half the guesses were kept, so it came out slower.
USB4 cabledirect box-to-box link
Head · llama-server :8081 · DSpark on
CC
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineROG Flow Z13 GZ302EA
- Memory128 GB unified
- Holdsits share + the 10.9 GB DSpark drafter + the bench
- Enginellama.cpp b1341 · llama-server
peak during run 179 °F
1 Gb networkhome router
Worker · RPC
Spark
NVIDIA GB10 Grace BlackwellCUDA 13ARMLinux
- MachineASUS Ascent GX10
- Memory128 GB unified
- Holdsits share of the model
- Enginellama.cpp b1341 · CUDA build
peak during run 135 °F
11.0tok/s
Answer speed, no history, with DSpark. The plain run did 14.5.
−32%worst
Biggest loss, on the 512 prompt after 4K history (14.5 → 9.8 tok/s)
51%kept
Draft guesses the big model accepted (1,258 of 2,450)
4/4clean
Every test finished, zero errors, zero heat alerts
-24%
512 prompt · no history
14.52 → 11.03 tok/s
-31%
2,048 prompt · no history
14.42 → 10.02 tok/s
-32%
512 prompt · +4K history
14.47 → 9.8 tok/s
-30%
2,048 prompt · +4K history
14.13 → 9.84 tok/s
Plain vs. DSpark, same three boxes, same test
The Hive Benchmarking Stack (llama-benchy 0.4.0): prompts of 512 and 2,048 tokens, 128-token answers, with no history and with 4,096 tokens of history. Each test runs 3 times. Both runs used the same model file, engine, boxes and flags; the only change is the DSpark drafter on the head.
Answer speed
tokens per second · higher is better
24–32% slower at every size. Half the guesses were rejected, and each round trip crosses the box-to-box links.
Prompt reading
tokens per second · higher is better
6–10% slower. The drafter also has to read the prompt.
Wait for the first word
seconds · lower is better
Up to 3 seconds longer, most on the long tests.
Plain 07310731 + DSpark, 3 guesses
Temperatures
CC carries the drafter as the head and peaked at 179 °F, the same as the plain run and well under the 203 °F automatic stop. Corsa peaked at 125 °F and Spark at 135 °F.
red line = 203 °F automatic stop
What this tells us
Speculative decoding only pays when guesses are cheap to check. On this cluster, GLM-5.3-Flash's built-in MTP layer (71% kept) made answers 23–37% faster, while DSpark (51% kept) made DeepSeek slower. Next we are testing DSpark on just the two AMD boxes over the faster USB4 cable.
Draft acceptance is summed over every logged request, including warm-ups. DSpark drafts up to 3 tokens per step (the recommended setting for this drafter).