community pool / model answer

Qwen3.5-9B

chat · Qwen/Qwen3.5-9B

47.1 tok/s · model median · 43 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B60 24GB	8-bit	30.4 tok/s	reported	366	- GB	4,096	llama.cpp
Arc Pro B60 24GB ×2	8-bit	30.4 tok/s	reported	367	- GB	4,096	llama.cpp
AMD Ryzen 7 8745H w/ Radeon 780M	4-bit	13.9 tok/s	reported	1,630	4 GB	32,768	llama.cpp
Intel(R) Core(TM) Ultra 7 155H	4-bit	8.5 tok/s	reported	-	- GB	2,048	llama.cpp
GTX 1080 Ti 11GB	4-bit	64.9 tok/s	reported	-	- GB	2,048	llama.cpp
GTX 1080 Ti 11GB	5-bit	34.3 tok/s	reported	-	9 GB	8,192	llama.cpp
Radeon 8060S Graphics 103GB	4-bit	255.8 tok/s	reported	124	7.7 GB	4,096	hipfire
Radeon AI Pro R9700 32GB	4-bit	371.8 tok/s	reported	63	7.7 GB	4,096	hipfire
Radeon AI Pro R9700 57GB	4-bit	15.8 tok/s	reported	1,791	- GB	2,048	llama.cpp
Radeon RX 5700 XT 8GB	4-bit	61.6 tok/s	reported	148	7.4 GB	4,096	hipfire
Radeon RX 6950 XT 16GB	4-bit	222 tok/s	reported	126	7.4 GB	4,096	hipfire
RTX 3060 12GB	3-bit	43 tok/s	reported	1,540	6.4 GB	4,608	llama.cpp
RTX 3060 12GB	4-bit	51.4 tok/s	measured	1,685	- GB	131,072	llama.cpp
RTX 3060 Ti 8GB	4-bit	63.5 tok/s	reported	249	- GB	2,048	llama.cpp
RTX 3080 Ti 12GB	8-bit	15.4 tok/s	reported	164,150	- GB	250,000	llama.cpp
RTX 3090 24GB ×2	8-bit	71.7 tok/s	reported	675	- GB	2,048	llama.cpp
RTX 3090 Ti 24GB	4-bit	119.1 tok/s	reported	156	8 GB	8,192	llama.cpp
RTX 5080 16GB	8-bit	79.6 tok/s	reported	82	- GB	4,096	llama.cpp
RTX 5090 32GB	6-bit	170 tok/s	reported	46	- GB	2,048	llama.cpp
RTX PRO 6000 Blackwell 96GB	16-bit	171.6 tok/s	reported	402	33.6 GB	8,192	vllm
RX 7900 XTX 24GB	4-bit	337 tok/s	measured	76	7.4 GB	4,096	hipfire
Ryzen AI Max 395 128GB	4-bit	37.2 tok/s	measured	-	6.4 GB	32,768	llama.cpp
Ultra 7 155H 32GB	4-bit	6.5 tok/s	reported	-	- GB	2,048	llama.cpp

Context tested

The table above is the tested context for each exact rig and quantization cell. Empty fields remain empty rather than being filled with an estimate.

back to The wall ->capture / correct ->