How much VRAM does GLM-4.7-Flash-heretic need?

Its published quantizations range from 9.61 GiB to 29.93 GiB in weights alone, before KV cache and runtime overhead.

GLM-4.7-Flash-heretic on Radeon RX 6500 XT

Can I run GLM-4.7-Flash-heretic on a Radeon RX 6500 XT?

Q: Can I run GLM-4.7-Flash-heretic on a Radeon RX 6500 XT?

No. No published quantization of GLM-4.7-Flash-heretic fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of GLM-4.7-Flash-heretic fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 9.61 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◐
Q8_0	29.93 GiB	31.0	31.1	31.3	31.7	32.6	34.4
Q5_1	29.75 GiB	30.8	30.9	31.1	31.5	32.4	34.2
Q4_K_M	24.06 GiB	25.1	25.2	25.4	25.9	26.7	28.5
Q6_K	23.27 GiB	24.3	24.4	24.6	25.1	25.9	27.7
Q5_K_M	20.15 GiB	21.2	21.3	21.5	21.9	22.8	24.6
Q5_K_S	19.59 GiB	20.6	20.7	20.9	21.4	22.3	24.0
Q4_1	17.86 GiB	18.9	19.0	19.2	19.7	20.5	22.3
Q4_K_S	16.25 GiB	17.3	17.4	17.6	18.0	18.9	20.7
IQ4_NL	16.14 GiB	17.2	17.3	17.5	17.9	18.8	20.6
IQ4_XS	15.28 GiB	16.3	16.4	16.6	17.1	17.9	19.7
IQ3_M	12.65 GiB	13.7	13.8	14.0	14.4	15.3	17.1
IQ2_M	9.61 GiB	10.6	10.7	11.0	11.4	12.3	14.0

●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, 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.