Worker · RPC :50053
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- Holds~84 GB of the model
- Enginellama.cpp b1328 · ggml-rpc-server
peak during run 145 °F
Hive Benchmarks · Oct 8, 2026
744 billion parameters. Three boxes. No API.
Full GLM-5.3, a frontier-class open model from Z.ai, ran as one model split across two AMD Strix Halo boxes and an NVIDIA Grace Blackwell box.
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~84 GB + the file + the bench
- Enginellama.cpp b1328 · llama-server
peak during run 179 °F
1 Gb networkSpark's ~64 GB share crossed here
Worker · RPC :50052
Spark
NVIDIA GB10 Grace BlackwellCUDA 13ARMLinux
- MachineASUS Ascent GX10
- Memory128 GB unified
- Holds~64 GB of the model
- Enginellama.cpp 8172e65 · CUDA build
peak during run 152 °F
744B params
Mixture of experts: 256 experts, 8 working on each word
7.2tok/s
Answer speed, no history. About half of DeepSeek.
52tok/s
Best prompt-reading speed (512-token prompt)
12/12clean
Every request finished, zero errors, zero heat alerts
GLM-5.3 vs. DeepSeek, 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. DeepSeek-V4-Flash ran on the same three boxes the same day.
Prompt reading
tokens per second · higher is better
About 40–50% of DeepSeek's speed. GLM-5.3 has three times the active parameters (40B vs 13B), so each token costs more work.
Answer speed
tokens per second · higher is better
About 7 words a second, steady across all four tests. Comfortable reading speed for a 744B model on home hardware.
Wait for the first word
seconds · lower is better
The weak spot. With 8 pages of history and a 4-page prompt, it reads for 2 min 15 s before the first word.
DeepSeek-V4-Flash · 87 GB · 3 boxesGLM-5.3 · 254 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 179 °F, the same as the DeepSeek run. No box came close to the cutoff.
red line = 203 °F automatic stop
Room on each box, measured with the model loaded
This time the boxes are close to full, on purpose. Spark carries less because it also runs other services.
AMD boxes: 96 GB is set aside for the GPU (the rest of the 128 GB runs Windows); the GLM share includes its working memory. Spark shares one 121 GB pool with Linux and its live services, so its free number is what's free for anything. Spark's disk is down to 46 GB because its worker keeps a local copy of every piece it receives, so the next load can skip the network.
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.
GLM-5.3 at 253.9 GB fills the 3-box pool to about 8 GB per AMD box, so this is the ceiling for three boxes. The 5-box rung assumes two more Corsairs joining over USB4.