← sources

Running SenseNova-Loop-9B nvfp4 looped transformer with 300-500 tok/s on a laptop 5090

source: reddit (reddit:LocalLLaMA)published 2026-11-18 02:18ingested 2026-11-19 02:39

retained text

Their architecture seems to translate quite well to existing kernels. They also have a small drafter. The speed numbers are aggregated across 8 parallel agents running. 256k context.

So far, I can't confirm it holds up well against Qwen 3.8 27B nvfp4, but I might have degraded it a bit, despite a calibration run. Here is a quick check of [SeraphimSerapis/tool-eval-bench](https://github.com/SeraphimSerapis/tool-eval-bench). Will do some more tests tomorrow, the 9B leaves a lot of vram free for FP8 tests.

**tool-eval-bench runs** (temperature 0, NVFP4 weights):

|Test|Qwen3.8 27B Swift T0|SenseNova-Loop-9B T0|SenseNova-Loop-9B T1|
|:-|:-|:-|:-|
|**Standard (69)**|**98**|90|91|
|**Hard mode (35)**|**87**|79|83|
|Error recovery|83|83|83|
|Context and state|100|75|85|
|Autonomous planning|67|67|83|
|Structured output|100|92|75|
|Safety|100|96|92|
|Hard: state across turns|0|50|83|
|Hard: tool failures|80|60|60|
|Hard: tenant separation|100|100|100|
|Hard: applying policy|100|100|100|
|Hard: injection|100|50|50|

FP8 KV, DFlash2, T0 = temperature 0, T1 = temperature 1.0, top_p 0.95

Hardware: Intel Core 9 290hx, 64gb ram, 1TB nvme5 SSD, rtx 5090 laptop with 24gb vram.

document fetched 2026-11-19 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xz3f4w/running_sensenovaloop9b_nvfp4_looped_transformer