How much VRAM does GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill need?

Its published quantizations range from 9.22 GiB to 55.79 GiB in weights alone, before KV cache and runtime overhead.

GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill on GeForce RTX 3080

Can I run GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill on a GeForce RTX 3080?

Q: Can I run GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill on a GeForce RTX 3080?

No. No published quantization of GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill fits a GeForce RTX 3080 at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill fits GeForce RTX 3080 at any context we compute, with q4_0 KV. The smallest shipped quantization is 9.22 GiB in weights alone, against 9.30 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— 10 GB card, 9.30 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◐
BF16	55.79 GiB	56.7	56.7	56.8	57.1	57.5	58.5
F16	55.79 GiB	56.7	56.7	56.8	57.1	57.5	58.5
Q8_0	29.66 GiB	30.5	30.6	30.7	30.9	31.4	32.3
Q6_K	22.92 GiB	23.8	23.8	24.0	24.2	24.7	25.6
Q5_K_M	19.80 GiB	20.7	20.7	20.8	21.1	21.5	22.5
Q4_K_M	16.89 GiB	17.8	17.8	17.9	18.2	18.6	19.6
IQ4_NL	15.79 GiB	16.7	16.7	16.8	17.1	17.5	18.5
IQ4_XS	14.93 GiB	15.8	15.9	16.0	16.2	16.7	17.6
Q3_K_M	13.39 GiB	14.3	14.3	14.4	14.7	15.1	16.1
IQ3_M	12.30 GiB	13.2	13.2	13.3	13.6	14.0	15.0
Q3_K_S	12.14 GiB	13.0	13.1	13.2	13.4	13.9	14.8
IQ3_XS	11.50 GiB	12.4	12.4	12.5	12.8	13.2	14.2
IQ2_M	9.22 GiB	10.1	10.1	10.3	10.5	11.0	11.9

●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.