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

Its published quantizations range from 10.68 GiB to 12.85 GiB in weights alone, before KV cache and runtime overhead.

Can I run gpt-oss-20b on a Radeon Pro W6400?

Q: Can I run gpt-oss-20b on a Radeon Pro W6400?

No. No published quantization of gpt-oss-20b fits a Radeon Pro W6400 at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of gpt-oss-20b fits Radeon Pro W6400 at any context we compute, with q4_0 KV. The smallest shipped quantization is 10.68 GiB in weights alone, against 3.72 GiB usable. CPU offload can still run it, slowly.

From the file· weights summed from filesFrom the file· KV computed per layerPredicted· speed and compute buffer

KV cache dtypef16 q8_0 q4_0— 4 GB card, 3.72 GiB usable after overhead

Every quantization at every context

total memory required; green fits, red does not

Quant	Weights●	4K◐	8K◐	16K◐	32K◐	64K◐	128K◐
F16	12.85 GiB	13.8	13.8	13.8	14.0	14.2	14.6
Q8_0	11.28 GiB	12.2	12.2	12.3	12.4	12.6	13.0
MXFP4	11.28 GiB	12.2	12.2	12.3	12.4	12.6	13.0
Q6_K	11.21 GiB	12.1	12.2	12.2	12.3	12.5	13.0
Q2_K_L	10.95 GiB	11.9	11.9	11.9	12.1	12.3	12.7
Q5_K_M	10.91 GiB	11.8	11.9	11.9	12.0	12.2	12.7
Q5_K_S	10.91 GiB	11.8	11.9	11.9	12.0	12.2	12.6
Q4_K_M	10.83 GiB	11.7	11.8	11.8	11.9	12.1	12.6
Q4_K_S	10.82 GiB	11.7	11.8	11.8	11.9	12.1	12.6
Q4_1	10.78 GiB	11.7	11.7	11.8	11.9	12.1	12.5
Q3_K_M	10.72 GiB	11.6	11.7	11.7	11.8	12.0	12.5
Q4_0	10.71 GiB	11.6	11.7	11.7	11.8	12.0	12.4
Q2_K	10.68 GiB	11.6	11.6	11.7	11.8	12.0	12.4
Q3_K_S	10.68 GiB	11.6	11.6	11.7	11.8	12.0	12.4

●From the file◐Predictedwhat these mean

Figures are GiB of total memory: weights plus KV cache plus compute buffer and backend overhead. Weights and KV are near-exact; the overhead term is modeled. Hover any cell for the breakdown.

Why other calculators disagree

A parameters × bits ÷ 8 estimate ignores two things that dominate at long context. First, the weights themselves are not the nominal rate — quantizations are mixtures, so the real file is consistently larger than the label implies. Second, most of this model's layers cache only a 128-token window rather than the full context.