Local model performance on your hardware

generation:

154.1 tokens/s

quant:

UD-Q4_K_S (GGUF)

User reports Qwen3.6-35B-A3B at 154.12 t/s with MTP on 2x AMD Radeon PRO W7800 48GB. Setup is llama.cpp (Lemonade stack) with UD-Q4_K_S GGUF and tensor-split across both cards; stock decode without MTP is 79.86 t/s, and MTP drops to 146.7 t/s at 8k and 132.6 t/s at 32k. Part of a registry PR adding 40 recipes for the W7800 48GB; other models in the table include Qwen3-Coder 30B-A3B, Ornith 1.5 35B-A3B, Gemma 4 26B-A4B-it, Qwen3.8 Uncensored and Qwen3.6-27B.

generation:

90.0 tokens/s

quant:

IQ3_XXS (GGUF)

kv:

Q8

User reports Qwen 3.8 Flash-Next at about 90 t/s on a Radeon Pro W7800 48GB with 64GB DDR5 system RAM. Setup is llama.cpp with an IQ3_XXS GGUF quant, 128k context and Q8 KV cache, run with default settings. The user is satisfied with the speed but the model fell into repetition loops after 3-4 minutes of reasoning on a 27KB JavaScript audit task, repeating the same steps dozens of times; they interrupted it twice and ask whether a reasoning budget setting is the cause.