How much VRAM does HunyuanOCR need?

Its published quantizations range from 0.24 GiB to 2.01 GiB in weights alone, before KV cache and runtime overhead.

Can I run HunyuanOCR on a Radeon RX 6500 XT?

Q: Can I run HunyuanOCR on a Radeon RX 6500 XT?

Yes. HunyuanOCR runs on a Radeon RX 6500 XT: the largest quantization that fits is F16 at 65,536 context with f16 KV, needing 3.57 GiB of 3.72 GiB usable memory.

Yes. The best fit is F16 at 65,536 context with f16 KV — 3.57 GiB of 3.72 GiB usable, leaving 0.15 GiB headroom. Expect roughly 19 tokens/sec (modeled, ±37%).

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◐
F32	2.01 GiB	3.0	3.1	3.3	3.7	4.6	6.3
F16	1.01 GiB	2.0	2.1	2.3	2.7	3.6	5.3
BF16	1.01 GiB	2.0	2.1	2.3	2.7	3.6	5.3
Q8_0	0.54 GiB	1.5	1.6	1.8	2.3	3.1	4.8
Q6_K	0.42 GiB	1.4	1.5	1.7	2.1	3.0	4.7
Q5_K_M	0.37 GiB	1.4	1.5	1.7	2.1	2.9	4.6
Q5_K_S	0.37 GiB	1.3	1.5	1.7	2.1	2.9	4.6
Q4_K_M	0.33 GiB	1.3	1.4	1.6	2.0	2.9	4.6
Q4_K_S	0.32 GiB	1.3	1.4	1.6	2.0	2.9	4.6
IQ4_XS	0.31 GiB	1.3	1.4	1.6	2.0	2.9	4.6
Q3_K_L	0.30 GiB	1.3	1.4	1.6	2.0	2.9	4.6
Q3_K_M	0.29 GiB	1.3	1.4	1.6	2.0	2.8	4.5
Q3_K_S	0.27 GiB	1.2	1.4	1.6	2.0	2.8	4.5
Q2_K	0.24 GiB	1.2	1.3	1.5	2.0	2.8	4.5

●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.

Run it

The best-fitting configuration above, as a command:

llama-cli -hf tencent/HunyuanOCR \
  --ctx-size 65536 \
  -ngl auto

Recent llama.cpp defaults to --fit on with -ngl auto, so it will size the offload for you. The question worth your attention is not how many layers to offload but what context and quantization you are willing to live with — which is what the grid above is for.

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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.