SenseNova-Loop 9B
on M5 Ultra 256GB · mlx-vlm · 1,024 ctx
Use cases
long-context
Summary
User benchmarks SenseNova-Loop-9B at 1K to 512K context on an M5 Ultra 256GB, comparing MLX and GGUF builds. The MLX 4-bit build generates 20.0 tok/s at 1K, 17.2 at 128K and 11.0 at 512K, reading the 512K prompt in 12.0 minutes with 71 GB peak memory.
MLX builds were converted with mlx-vlm; GGUF builds ran on the OpenSenseNova llama.cpp fork. MLX reads prompts at 2,200 to 2,500 tok/s at 32K against about 1,000 for GGUF.
GGUF holds about 6 to 8 tok/s at any length on roughly half the memory; Q4_K_M with an 8-bit KV cache peaked at 15 GB at 256K. All five builds retrieved all three hidden codes at every length, 512K included. Mainline llama.cpp and LM Studio do not load these GGUFs yet.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Models most run on the M5 Ultra 256GB
- 73.4 t/s · one run
- Gemma 4 31B16-bit13.9 t/s · one run
- 74.7 t/s · one run
- 17.0 t/s · one run
- 141.0 t/s · one run
- Qwen3.8 27B4-bit50.0 t/s · one run
Bars are medians on one scale, 0 to 141.0 tokens/s. Press ▶ to watch text arrive at a row's speed.