OpenAI · 21B · MXFP4 · 11.3 GiB

gpt-oss-20b at MXFP4 needs about 12.6 GiB at 8K context, and the smallest listed machine that fits it is GeForce RTX 4060 Ti 16GB.

Open-weight reasoning MoE. The published GGUF is MXFP4, not Q4_K_M. 4 of 32 experts run each token. Half the layers use a 128-token sliding window.

ollama run gpt-oss:20b

Weights: gpt-oss-20b-MXFP4.gguf. Architecture: config source. Hugging Face repo created 2025-08-04. License field: apache-2.0.

Quants with a measured file size

QuantFileSize
MXFP4 · recommendedgpt-oss-20b-MXFP4.gguf11.3 GiB

Machines

Same quant and 8,192 tokens on every row.
MachineMemoryTotalVerdictSpeed
Apple M1, 8 GB8 GB unified12.6 GiBNo—
Apple M2, 8 GB8 GB unified12.6 GiBNo—
Apple M3, 8 GB8 GB unified12.6 GiBNo—
GeForce RTX 3060 Ti8 GB GDDR12.6 GiBOffload14–20 tok/s
GeForce RTX 30708 GB GDDR12.6 GiBOffload14–20 tok/s
GeForce RTX 40608 GB GDDR12.6 GiBOffload13–19 tok/s
GeForce RTX 4060 Ti 8GB8 GB GDDR12.6 GiBOffload13–19 tok/s
GeForce RTX 50608 GB GDDR12.6 GiBOffload25–37 tok/s
GeForce RTX 5060 Ti 8GB8 GB GDDR12.6 GiBOffload25–37 tok/s
Radeon RX 66008 GB GDDR612.6 GiBOffload13–18 tok/s
Radeon RX 76008 GB GDDR612.6 GiBOffload13–19 tok/s
GeForce RTX 3080 10GB10 GB GDDR12.6 GiBOffload26–38 tok/s
GeForce RTX 3060 12GB12 GB GDDR12.6 GiBOffload75–109 tok/s
GeForce RTX 407012 GB GDDR12.6 GiBOffload105–153 tok/s
GeForce RTX 4070 Super12 GB GDDR12.6 GiBOffload105–153 tok/s
GeForce RTX 4070 Ti12 GB GDDR12.6 GiBOffload105–153 tok/s
GeForce RTX 507012 GB GDDR12.6 GiBOffload140–204 tok/s
Intel Arc B58012 GB GDDR612.6 GiBOffload95–138 tok/s
Radeon RX 6700 XT12 GB GDDR612.6 GiBOffload80–116 tok/s
Radeon RX 7700 XT12 GB GDDR612.6 GiBOffload90–131 tok/s
Apple M1 Pro, 16 GB16 GB unified12.6 GiBNo—
Apple M1, 16 GB16 GB unified12.6 GiBNo—
Apple M2 Pro, 16 GB16 GB unified12.6 GiBNo—
Apple M2, 16 GB16 GB unified12.6 GiBNo—
Apple M3, 16 GB16 GB unified12.6 GiBNo—
Apple M4, 16 GB16 GB unified12.6 GiBNo—
GeForce RTX 4060 Ti 16GB16 GB GDDR12.6 GiBFits60–95 tok/s
GeForce RTX 4070 Ti Super16 GB GDDR12.6 GiBFits141–222 tok/s
GeForce RTX 408016 GB GDDR12.6 GiBFits150–237 tok/s
GeForce RTX 4080 Super16 GB GDDR12.6 GiBFits154–243 tok/s
GeForce RTX 5060 Ti 16GB16 GB GDDR12.6 GiBFits94–148 tok/s
GeForce RTX 5070 Ti16 GB GDDR12.6 GiBFits188–296 tok/s
GeForce RTX 508016 GB GDDR12.6 GiBFits201–317 tok/s
Intel Arc A770 16GB16 GB GDDR612.6 GiBFits107–169 tok/s
Legion Go16 GB LPDDR5X-750012.6 GiBNo—
Radeon RX 6800 XT16 GB GDDR612.6 GiBFits107–169 tok/s
Radeon RX 7800 XT16 GB GDDR612.6 GiBFits131–206 tok/s
Radeon RX 9060 XT 16GB16 GB GDDR612.6 GiBFits67–106 tok/s
Radeon RX 9070 XT16 GB GDDR612.6 GiBFits134–211 tok/s
ROG Ally16 GB LPDDR5-640012.6 GiBNo—
Steam Deck LCD16 GB LPDDR5-550012.6 GiBNo—
Steam Deck OLED16 GB LPDDR5-640012.6 GiBNo—
Apple M3 Pro, 18 GB18 GB unified12.6 GiBTight31–50 tok/s
Radeon RX 7900 XT20 GB GDDR612.6 GiBFits167–264 tok/s
Apple M2, 24 GB24 GB unified12.6 GiBFits21–33 tok/s
Apple M3, 24 GB24 GB unified12.6 GiBFits21–33 tok/s
Apple M4 Pro, 24 GB24 GB unified12.6 GiBFits57–90 tok/s
Apple M4, 24 GB24 GB unified12.6 GiBFits25–40 tok/s
GeForce RTX 309024 GB GDDR12.6 GiBFits196–309 tok/s
GeForce RTX 409024 GB GDDR12.6 GiBFits211–333 tok/s
Radeon RX 7900 XTX24 GB GDDR612.6 GiBFits201–317 tok/s
ROG Ally X24 GB LPDDR5X-750012.6 GiBFits25–40 tok/s
Apple M1 Max, 32 GB32 GB unified12.6 GiBFits84–132 tok/s
Apple M1 Pro, 32 GB32 GB unified12.6 GiBFits42–66 tok/s
Apple M2 Max, 32 GB32 GB unified12.6 GiBFits84–132 tok/s
Apple M2 Pro, 32 GB32 GB unified12.6 GiBFits42–66 tok/s
Apple M4, 32 GB32 GB unified12.6 GiBFits25–40 tok/s
GeForce RTX 509032 GB GDDR12.6 GiBFits375–592 tok/s
Apple M3 Pro, 36 GB36 GB unified12.6 GiBFits31–50 tok/s
Apple M4 Max 14-core, 36 GB36 GB unified12.6 GiBFits86–135 tok/s
Apple M3 Max 16-core, 48 GB48 GB unified12.6 GiBFits84–132 tok/s
Apple M4 Max 16-core, 48 GB48 GB unified12.6 GiBFits114–180 tok/s
Apple M4 Pro, 48 GB48 GB unified12.6 GiBFits57–90 tok/s
Apple M1 Max, 64 GB64 GB unified12.6 GiBFits84–132 tok/s
Apple M2 Max, 64 GB64 GB unified12.6 GiBFits84–132 tok/s
Apple M3 Max 16-core, 64 GB64 GB unified12.6 GiBFits84–132 tok/s
Apple M4 Max 16-core, 64 GB64 GB unified12.6 GiBFits114–180 tok/s
Apple M4 Pro, 64 GB64 GB unified12.6 GiBFits57–90 tok/s
Apple M2 Max, 96 GB96 GB unified12.6 GiBFits84–132 tok/s
Apple M3 Max 16-core, 128 GB128 GB unified12.6 GiBFits84–132 tok/s
Apple M4 Max 16-core, 128 GB128 GB unified12.6 GiBFits114–180 tok/s

Exact answers

GeForce RTX 3060 12GBoffloadGeForce RTX 4060offloadGeForce RTX 4060 Ti 16GBfitsGeForce RTX 4070offloadGeForce RTX 4070 Ti SuperfitsGeForce RTX 4090fitsGeForce RTX 5070offloadGeForce RTX 5080fits

FAQ

Why MXFP4?

MXFP4 is the quant this page recommends. The file is gpt-oss-20b-MXFP4.gguf at 11.3 GiB. Other single-file quants are in the table when a real size was available.

Where do the architecture numbers come from?

num_local_experts 32, num_experts_per_tok 4, intermediate_size 2880, sliding_window 128, layer_types. Source: https://huggingface.co/openai/gpt-oss-20b/raw/main/config.json

Does this include a long context?

The fit uses 8,192 tokens, or less when that is all the fast memory can hold. The model file allows up to 131,072 tokens, and KV cache grows as you use more of that.

What is the Ollama command?

ollama run gpt-oss:20b. The library tag is Q4-class or the vendor quant. Compare it with the GGUF linked on this page before you treat the sizes as identical.