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 131 °F
Hive Benchmarks · Oct 11, 2026
Same model. Same three boxes. New engine reads up to 30% faster.
DeepSeek-V4-Flash, split across two AMD boxes and one NVIDIA box. We reran the exact earlier 3-box test with one change: the llama.cpp engine went from b1328 to b1341. Prompt reading got 14–30% faster and the wait for the first word dropped by up to 11 seconds. Answer speed moved only 2–4%.
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 140 °F
158tok/s
Best prompt reading (512 prompt after 4K history). The b1328 run did 126.
−11sec
Shorter wait for the first word on the biggest test (51.0 → 39.9 s)
13.95tok/s
Answer speed, no history. The b1328 run did 13.4, a small 2–4% gain.
4/4clean
Every test finished, zero errors, zero heat alerts
+30%
512 prompt · no history
98.3 → 128.1 tok/s
+14%
2,048 prompt · no history
126.7 → 144.6 tok/s
+26%
512 prompt · +4K history
125.7 → 157.9 tok/s
+28%
2,048 prompt · +4K history
120.6 → 153.9 tok/s
Old engine vs. new engine, 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 file, boxes and flags; the only change is the engine build.
Prompt reading
tokens per second · higher is better
14–30% faster at every size. This is where the new engine helps most.
Wait for the first word
seconds · lower is better
12–23% shorter. Faster prompt reading shows up directly as less waiting.
Answer speed
tokens per second · higher is better
Up 2–4%. Answer speed is held back by the links between boxes, not by the engine.
Engine b1328Engine b1341
Temperatures
All three boxes were logged. CC does the most work as the head and peaked at 179 °F, the same as the b1328 run and well under the 203 °F automatic stop. Corsa peaked at 131 °F and Spark at 140 °F.
red line = 203 °F automatic stop
Where this sits on the ladder
This model fits on two boxes; the third box adds room, not speed. Bar widths are pooled model memory; dashed lines are each model's file size.
Next speed lever: a faster link for Spark (1 Gb now, 10 Gb possible) for answer speed and the first-word wait.