community pool / model answer

gpt-oss-20b

chat · openai/gpt-oss-20b

102.8 tok/s · model median · 39 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B60 24GB	8-bit	40.7 tok/s	reported	459	- GB	4,096	llama.cpp
Arc Pro B60 24GB ×2	8-bit	40.1 tok/s	reported	460	- GB	4,096	llama.cpp
GB10 Grace Blackwell 121GB	16-bit	49.1 tok/s	reported	-	- GB	128,000	vllm
GB10 Grace Blackwell 128GB	4-bit	49.2 tok/s	reported	74	- GB	32,768	vllm
M1 Max 64GB	4-bit	62.4 tok/s	reported	172	- GB	2,048	ollama
M1 Max 64GB	8-bit	83.6 tok/s	reported	-	- GB	640	llama.cpp
M5 Max 128GB	4-bit	137 tok/s	reported	-	12.4 GB	2,048	mlx
Radeon 890M 32GB	4-bit	23.9 tok/s	reported	1,210	- GB	2,048	llama.cpp
Radeon AI Pro R9700 32GB ×2	4-bit	48.3 tok/s	reported	89	- GB	32,768	vllm
Radeon AI Pro R9700 32GB ×3	8-bit	152.9 tok/s	measured	246	17 GB	131,072	llama.cpp
RTX 2080 8GB	4-bit	12 tok/s	reported	-	12 GB	2,048	ollama
RTX 3060 12GB	4-bit	74 tok/s	reported	-	12 GB	2,048	ollama
RTX 3080 12GB	4-bit	21 tok/s	reported	-	8.2 GB	2,048	ollama
RTX 3090 24GB	4-bit	162.4 tok/s	measured	1,790	0 GB	131,072	llama.cpp
RTX 5080 16GB	8-bit	222.4 tok/s	reported	48	- GB	4,096	llama.cpp
RTX 5090 32GB	4-bit	307.2 tok/s	reported	83	- GB	2,048	lmstudio
RTX3060+RTX2080-pooled-20GB 20GB	4-bit	77 tok/s	reported	-	12 GB	2,048	ollama
RX 6700 XT 12GB	8-bit	22.4 tok/s	measured	-	- GB	4,736	llama.cpp
Ryzen AI Max 385 24GB	4-bit	23.2 tok/s	reported	1,210	- GB	2,048	llama.cpp
Ryzen AI Max 395 64GB	4-bit	75.2 tok/s	reported	931	12.2 GB	8,192	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.