SenseNova-U1.5 8B MoT
M5 Ultra · llama.cpp
- generation:
- 125.0 tokens/s
- prompt processing (prefill):
- 3482.0 tokens/s
- quant:
- Q4_K_M (GGUF)
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
vision
User reports SenseNova-U1.5-8B-MoT at 125 t/s generation and 3482 t/s prompt processing on an M5 Ultra. Setup is llama.cpp (llama-bench, Metal, all layers on GPU) with the Q4_K_M GGUF at 5.32 GiB; BF16 gave 56 t/s and Q8_0 88 t/s on the same table. Through llama-mtmd-cli the Q4_K_M model caught a deliberately wrong $1.00 subtotal on a fake receipt in 16 s including load, and read back a menu, a whiteboard and three bilingual signs without error in 4 to 10 s each, though those images were made by the same model.