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 129 °F
Hive Benchmarks · Oct 11, 2026
The newer DeepSeek. Our fastest DeepSeek on three boxes.
DeepSeek-V4-Flash-0731 is the July refresh of DeepSeek-V4-Flash, same size and design. We ran it on the same three boxes, engine and test as the original. It answered 4–7% faster and read prompts 4–7% faster.
USB4 cabledirect box-to-box link
Head · llama-server :8081
CC
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineROG Flow Z13 GZ302EA
- Memory128 GB unified
- Holdsits share + the model file + 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 136 °F
14.5tok/s
Answer speed, no history. The original V4-Flash did 13.95 on the same setup.
169tok/s
Best prompt reading (512 prompt after 4K history)
3.9sec
Wait for the first word on a 512-token prompt
4/4clean
Every test finished, zero errors, zero heat alerts
+4%
512 prompt · no history
13.95 → 14.52 tok/s
+6%
2,048 prompt · no history
13.57 → 14.42 tok/s
+7%
512 prompt · +4K history
13.55 → 14.47 tok/s
+6%
2,048 prompt · +4K history
13.39 → 14.13 tok/s
Original V4-Flash vs. 0731, same three boxes, same engine, 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. Same engine (b1341), boxes and flags; the change is the model release and its quant (original IQ2XXS 86.7 GB vs. unsloth UD-IQ2_XXS 90.9 GB).
Answer speed
tokens per second · higher is better
4–7% faster at every size. Holds above 14 tok/s even with 4K of history.
Prompt reading
tokens per second · higher is better
4–7% faster.
Wait for the first word
seconds · lower is better
Slightly shorter at every size.
DeepSeek-V4-Flash (original)DeepSeek-V4-Flash-0731
Temperatures
CC as the head peaked at 179 °F, the same as the original model's run and well under the 203 °F automatic stop. Corsa peaked at 129 °F and Spark at 136 °F.
red line = 203 °F automatic stop
Where this sits on the ladder
0731 is the version DeepSeek's own speed helper (DSpark) was made for. We tested that helper next on this same setup; see its own page.
Next speed lever: a faster link for Spark (1 Gb now, 10 Gb possible).