Worker · RPC :50053
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- Holds~70 GB of the model
- Enginellama.cpp b1341 · ggml-rpc-server
peak during run not logged
Hive Benchmarks · Oct 9, 2026
Same model. Same three boxes. Up to 37% faster.
GLM-5.3-Flash ships with a small built-in helper layer that guesses the next words ahead of time (multi-token prediction, MTP). The big model checks two guesses at once, so it keeps more words per pass with no change in answer quality. We switched it on and reran the exact same test as the plain run.
USB4 cabledirect box-to-box link
Head · llama-server :8081 · MTP on
CC
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineROG Flow Z13 GZ302EA
- Memory128 GB unified
- Holds~74 GB + the MTP helper + the bench
- Enginellama.cpp b1341 · llama-server
peak during run 183 °F
1 Gb networkunchanged from the plain run
Worker · RPC :50052
Spark
NVIDIA GB10 Grace BlackwellCUDA 13ARMLinux
- MachineASUS Ascent GX10
- Memory128 GB unified
- Holds~53 GB of the model
- Enginellama.cpp b86d2f0 (= b1341) · CUDA build
peak during run 144 °F
14.2tok/s
Answer speed, no history. The plain run did 10.8 on the same boxes.
+37%best
Biggest gain, on the 2,048-token prompt (10.2 → 14.0 tok/s)
71%kept
Helper guesses the big model accepted (1,208 of 1,708)
12/12clean
Every request finished, zero errors, zero heat alerts
+32%
512 prompt · no history
10.8 → 14.2 tok/s
+37%
2,048 prompt · no history
10.2 → 14.0 tok/s
+30%
512 prompt · +4K history
10.1 → 13.2 tok/s
+23%
2,048 prompt · +4K history
10.0 → 12.4 tok/s
Plain vs. MTP, 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, engine, boxes and split; the only change is the MTP helper.
Answer speed
tokens per second · higher is better
23–37% faster at every size. This is the number you feel: words appear about a third quicker.
Prompt reading
tokens per second · higher is better
Flat, within 1–4%. The helper only works on the answer, not on reading the prompt, so this is what we expected.
Wait for the first word
seconds · lower is better
About the same, 2–4% longer. Prompt reading still sets the wait; the faster network link is the lever for this one.
PlainMTP on, 2 guesses
Temperatures
CC does the most work as the head and now runs the helper too. It peaked at 183 °F, 5 °F above the plain run and well under the 203 °F automatic stop. Spark matched the plain run at 144 °F. Corsa's logger stopped right after it started, so Corsa has no reading for this run; its only reading, at the start, was 109 °F.
red line = 203 °F automatic stop
Room on each box, measured with MTP loaded
The helper costs memory on the head. CC went from ~69 GB to ~74 GB; the workers barely moved. There's still real room on every box.
AMD boxes: 96 GB is set aside for the GPU (the rest of the 128 GB runs Windows). Spark shares one 121 GB pool with Linux and its live services, so its free number is what's free for anything.
Where this sits on the ladder
Adding boxes adds room. The MTP helper adds speed without adding a box. Bar widths are pooled model memory; dashed lines are each model's file size.
Next speed levers: a faster link for Spark (1 Gb now, 10 Gb possible) for the first-word wait, and trying 1 or 3 guesses per step as separate recorded runs. MTP acceptance is summed over every logged request, including warm-ups. Load: tensor split 75,50,75 in load order (Corsa, Spark, CC), 32K context, head started with -lm mmap.