Hive Cluster Benchmarks

Frontier models. Home hardware. No API.

One model split across two AMD Strix Halo boxes and an NVIDIA Grace Blackwell box with llama.cpp RPC. x86 and ARM, ROCm and CUDA, Windows and Linux.

DeepSeek-V4-Flash-0731 + DSpark · 2 boxes

284B MoE · official draft model on · 102 GB · 2 boxes

20.1tok/s answers

+37%vs plain

48%guesses kept

Open the benchmark →

DeepSeek-V4-Flash-0731 · 2 boxes

284B MoE · UD-IQ2_XXS · 91 GB · 2 boxes

15.0tok/s answers

259tok/s prompt

+53%vs 3 boxes

Open the benchmark →

DeepSeek-V4-Flash-0731 + DSpark

284B MoE · official draft model on · 102 GB · 3 boxes

11.0tok/s answers

−24%vs plain

51%guesses kept

Open the benchmark →

DeepSeek-V4-Flash-0731

284B MoE · UD-IQ2_XXS · 91 GB · 3 boxes

14.5tok/s answers

169tok/s prompt

4/4clean

Open the benchmark →

DeepSeek-V4-Flash · new engine

284B MoE · IQ2XXS · 87 GB · b1328 → b1341

13.95tok/s answers

+30%prompt reading

4/4clean

Open the benchmark →

GLM-5.3-Flash + MTP

320B MoE · built-in draft layer on · 200 GB · 3 boxes

14.2tok/s answers

+37%best gain

12/12clean

Open the benchmark →

GLM-5.3-Flash

320B MoE · UD-Q4_K_XL · 200 GB · 3 boxes

10.8tok/s answers

137tok/s prompt

12/12clean

Open the benchmark →

GLM-5.3

744B MoE · UD-Q2_K_XL · 254 GB · 3 boxes

7.2tok/s answers

52tok/s prompt

12/12clean

Open the benchmark →

DeepSeek-V4-Flash

284B MoE · IQ2XXS · 87 GB · 2 vs 3 boxes

13.4tok/s answers

127tok/s prompt

~290GB pooled

Open the benchmark →