Best home rig for gpt-oss 20B

Will it fit my GPU? Best value GPU per euro →

Real, measured generation speeds for gpt-oss 20B on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s	Cost
RTX 5090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 324 tok/s, Qwen3.8 27B 75 tok/s	reported as Q4_K_M	llama.cpp 0.7.0-dev	324.2		source
RTX 5090: 2026 models (llamabench community)	llama-bench tg128 (quant per result)	llama.cpp	313		source
RTX Pro 6000 96GB workstation card	MXFP4	llama.cpp	286.91		source
Single RTX 5090 32GB	MXFP4	llama.cpp	282.51		source
Intel Core Ultra 265K + RTX 4090		llama.cpp	279.65		source
RTX 3090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 207 tok/s, Qwen3.6 27B 43 tok/s	reported as Q4_K_M	llama.cpp 0.7.0-dev	207.2		source
Single RTX 5080 16GB	MXFP4	llama.cpp	204.85		source
RTX 4090: 2026 models (llamabench community)	llama-bench tg128 (quant per result)	llama.cpp	203		source
Single RTX 5070 Ti 16GB	MXFP4	llama.cpp	189.45		source
Intel Core Ultra 265K + Radeon AI Pro R9700		llama.cpp	163.26		source
Intel Core Ultra 265K + RTX 4070 Ti		llama.cpp	163.14		source
Single RTX 3090 24GB	MXFP4	llama.cpp	161.77		source
Mac Studio M2 Ultra 76-core GPU, 192 GB	MXFP4	llama.cpp	116.08		source
Mac Studio M3 Ultra 512GB (Geerling test)	MXFP4	llama.cpp (Metal, llama-bench)	115.52		source
Mac Studio M3 Ultra 512GB (Geerling test)		llama.cpp	115.29		source
Single RTX 5060 Ti 16GB	MXFP4	llama.cpp	111.51		source
Mac Studio M1 Ultra (64-core GPU) 128GB		llama.cpp	100.43		source
DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s		vLLM 0.31.0, 128K ctx, 1 user	96.2		source
Intel Core Ultra 265K + RTX A4000		llama.cpp	93.55		source
Apple M4 Max 36GB	MXFP4	llama.cpp	92.36		source
Dell Pro Max with GB10 (Nvidia DGX Spark)		llama.cpp	87.97		source
Intel Core Ultra 265K + RTX 3060		llama.cpp	83.96		source
Apple M1 Max 64GB	MXFP4	llama.cpp	75.15		source
HP OmniBook Ultra 16 (RTX Spark N1X, 128GB) laptop: local AI benchmarks	Ollama default, 64K context, Codex CLI coding agent	Ollama	58.0	$6,230 (≈€5,559)	source

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Compare: DGX Spark vs Mac Studio vs Strix Halo: measured tok/s and tok/s per watt

Memory math: how much VRAM / RAM does gpt-oss 20B need?

gpt-oss 20B has 20.91B parameters, about 3.6B active per token (source: GGUF param count via llamabench.ai (OpenAI model card: 21B total, 3.6B active)). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.

Format	Bits/weight	Weights	Smallest common memory size that holds the weights alone
Native release format	–	≈13 GB	16 GB
Q8_0	8.50	22.2 GB	24 GB
Q6_K	6.56	17.2 GB	24 GB
Q5_K_M	5.70	14.9 GB	16 GB
Q4_K_M	4.89	12.8 GB	16 GB
Q3_K_M	4.00	10.4 GB	12 GB
Q2_K	3.16	8.3 GB	12 GB

Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.

Why it can still run on smaller GPUs: it is a mixture-of-experts model, so only about 3.6B of its 20.91B parameters are used for each generated token. Runtimes such as llama.cpp (--n-cpu-moe), ExLlamaV3 or Project Maya can keep most expert weights in system RAM and only the shared layers and hot experts in VRAM. Then you need VRAM + system RAM ≥ weights, and speed is limited mainly by RAM bandwidth.

Measured on setups with less GPU memory than the Q4_K_M / native weights (so part of the model lived in system RAM or was streamed from disk):

RTX 4070 Ti (12 GB): 163.14 tok/s ()
RTX 3060 (12 GB): 83.96 tok/s ()

FAQ

How much VRAM does gpt-oss 20B need?

The weights are about 13 GB in its native format (GGUF param count via llamabench.ai (OpenAI model card: 21B total, 3.6B active)). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is RTX 5080 (16 GB) at 204.85 tok/s.

What is the fastest measured hardware for gpt-oss 20B?

RTX 5090: 324.2 tok/s generation (llama.cpp 0.7.0-dev, reported as Q4_K_M), from https://llm-speed.com/r/r_dwp5fgdapmw.

Which GPUs have measured results for gpt-oss 20B?

Apple M1 Max, Apple M1 Ultra, Apple M2 Ultra, Apple M3 Ultra, Apple M4 Max, DGX Spark (GB10), RTX 3060 12GB, RTX 3090, RTX 4070 Ti, RTX 4090, RTX 5060 Ti 16GB, RTX 5070 Ti, RTX 5080, RTX 5090, RTX A4000, RTX Pro 6000 (Blackwell, 96 GB), RTX Spark N1X, Radeon AI Pro R9700.

Can it run on… Apple M1 Max · Apple M1 Ultra · Apple M2 Ultra · Apple M3 Ultra · Apple M4 Max · DGX Spark (GB10) · RTX 3060 12GB · RTX 3090 · RTX 4070 Ti · RTX 4090 · RTX 5060 Ti 16GB · RTX 5070 Ti · RTX 5080 · RTX 5090 · RTX A4000 · RTX Pro 6000 (Blackwell, 96 GB) · RTX Spark N1X · Radeon AI Pro R9700