gpt-oss 20B on the NVIDIA RTX 4060

Best measured: 42.3 tokens/s generation and 1,642 tokens/s prompt processing (MXFP4, CUDA, FA on).

For comparison, our size-scaling estimate for this combination is 45.4 tok/s.

Every run

Quant	Backend	Generate tok/s	Prompt tok/s	Test	Build / by / notes	Source
MXFP4	CUDA FA on	42.3	1,642	tg128	Djip007 llama-bench model column: gpt-oss 20B MXFP4 MoE; model file is larger than this card's VRAM, so part of it ran on the CPU/system RAM	llama.cpp discussion #15396
F16	CUDA FA on	32.7	1,430	tg128	Djip007 llama-bench model column: gpt-oss 20B F16; model file is larger than this card's VRAM, so part of it ran on the CPU/system RAM	llama.cpp discussion #15396

tg128 = generating 128 tokens; pp512 = reading a 512-token prompt (llama-bench). LocalScore averages several prompt and output lengths, so its numbers run lower than tg128.

Check prices

Amazon Newegg

As an Amazon Associate I earn from qualifying purchases.

Doesn't fit in this card's VRAM

Part of this model ran on the CPU here. A rented 48-80 GB GPU keeps it all in VRAM.

Rent a bigger GPU on RunPod

Affiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.

All results for the NVIDIA RTX 4060 · gpt-oss 20B on other hardware