How much VRAM does gpt-oss-20b need?

Q4_K_M is exactly 11,624,759,488 bytes (10.83 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is gpt-oss-20b's KV cache?

0.77 GiB at 32K context with an f16 cache, computed per layer from the model's architecture — most of its layers use sliding-window attention, so a flat formula overstates this substantially. Quantizing the cache to q8_0 roughly halves it.

Is gpt-oss-20b a mixture-of-experts model?

Yes — 32 experts with 4 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

openai · text · mixture of experts

gpt-oss-20b

openai/gpt-oss-20b

gpt-oss-20b at Q4_K_M is exactly 11,624,759,488 bytes (10.83 GiB / 11.62 GB) — an effective 4.323 bits per weight, not the nominal 4. Its KV cache at 32K is 0.77 GiB, not the 1.50 GiB a flat formula predicts.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

21.5B

total, not active

Architecture

gpt-oss

24 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q3_K_S	10.68 GiB	11,463,894,208	4.263	—	unsloth
Q2_K	10.68 GiB	11,468,317,888	4.265	—	unsloth
Q4_0	10.71 GiB	11,501,495,488	4.277	459	unsloth
Q3_K_M	10.72 GiB	11,506,103,488	4.279	459	unsloth
Q4_1	10.78 GiB	11,577,504,448	4.306	—	unsloth
Q4_K_S	10.82 GiB	11,618,492,608	4.321	—	unsloth
Q4_K_M	10.83 GiB	11,624,759,488	4.323	459	unsloth
Q5_K_S	10.91 GiB	11,711,827,648	4.356	—	unsloth
Q5_K_M	10.91 GiB	11,717,357,248	4.357	459	unsloth
Q2_K_L	10.95 GiB	11,757,884,608	4.373	—	unsloth
Q6_K	11.21 GiB	12,041,000,128	4.478	459	unsloth
MXFP4	11.28 GiB	12,109,565,632	4.503	—	lmstudio-community
MXFP4	11.28 GiB	12,109,566,624	4.503	—	ggml-org
Q8_0	11.28 GiB	12,109,567,168	4.503	459	unsloth
F16	12.85 GiB	13,792,639,168	5.129	459	unsloth

KV cache by context

computed per layer — this model uses sliding-window attention

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.11 GiB	0.19 GiB	1.68×	12 / 12 / 0
8,192	0.21 GiB	0.38 GiB	1.83×	12 / 12 / 0
16,384	0.39 GiB	0.75 GiB	1.91×	12 / 12 / 0
32,768	0.77 GiB	1.50 GiB	1.95×	12 / 12 / 0
65,536	1.52 GiB	3.00 GiB	1.98×	12 / 12 / 0
131,072	3.02 GiB	6.00 GiB	1.99×	12 / 12 / 0

12 of 24 layers cache only a 128-token window rather than the full context, on a period of 2. Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 11.27 GiB. The real file is 10.83 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 1.50 GiB at 32K context where the real figure is 0.77 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

Hidden size

2880

Vocab

201,088

Sliding window

128

SWA period

MLA

Experts

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does gpt-oss-20b need?: Q4_K_M is exactly 11,624,759,488 bytes (10.83 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is gpt-oss-20b's KV cache?: 0.77 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is gpt-oss-20b a mixture-of-experts model?: Yes — 32 experts, 4 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of gpt-oss-20b should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.