Worker · RPC :50053
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- Holds~68 GB of the model
- Enginellama.cpp b1341 · ggml-rpc-server
peak during run 122 °F
Hive Benchmarks · Oct 9, 2026
320 billion parameters. Three boxes. Ten words a second.
GLM-5.3-Flash, the fast sibling of Z.ai's GLM-5.3, ran as one model split across two AMD Strix Halo boxes and an NVIDIA Grace Blackwell box. It's 200 GB on disk at 4-bit, too big for two boxes, and it answers about 45% faster than full GLM-5.3 on the same boxes.
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
- Holds~69 GB + the file + the bench
- Enginellama.cpp b1341 · llama-server
peak during run 178 °F
1 Gb networkSpark's ~53 GB share crossed here
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
320B params
Mixture of experts: 288 experts, 8 plus 1 shared working on each word
10.8tok/s
Answer speed, no history. GLM-5.3 did 7.2 on the same boxes.
137tok/s
Best prompt-reading speed, about 2.6× GLM-5.3's best
12/12clean
Every request finished, zero errors, zero heat alerts
GLM-5.3-Flash vs. full GLM-5.3, 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. Full GLM-5.3 ran on the same three boxes the day before.
Prompt reading
tokens per second · higher is better
1.7× to 2.9× faster. Flash puts ~18B parameters to work per token against GLM-5.3's 40B, spread over 45 layers instead of 78.
Answer speed
tokens per second · higher is better
About 10 words a second, steady across all four tests. That's 40–50% faster than GLM-5.3 at every size.
Wait for the first word
seconds · lower is better
Worst case drops from 2 min 15 s to 46 s. Still the slowest part, but now usable for real work.
GLM-5.3 · 254 GB · 3 boxesGLM-5.3-Flash · 200 GB · 3 boxes
Temperatures, logged the whole run
Every box ran its own logger with a 203 °F cutoff that stops the work on its own. CC works hardest as the head and peaked at 178 °F, 1 °F under the GLM-5.3 run. No box came close to the cutoff.
red line = 203 °F automatic stop
Room on each box, measured with the model loaded
Flash leaves real breathing room: about 27 GB free on each AMD box and about 51 GB on Spark.
AMD boxes: 96 GB is set aside for the GPU (the rest of the 128 GB runs Windows); the Flash share includes its working memory for a 32K context. 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, not speed. Room decides which models can live here without an API. Bar widths are pooled model memory; dashed lines are each model's file size.
At 199.7 GB, Flash is too big for the two-box pool (~192 GB) and fits the three-box pool with ~90 GB to spare. The file also carries a built-in next-word draft layer (multi-token prediction) that this run did not use. Model facts from the GGUF header: 288 experts, 8 used + 1 shared per token, 45 layers + 1 draft layer. Load: tensor split 75,50,75 in load order (Corsa, Spark, CC), 32K context, head started with -lm mmap.