How much VRAM does Kimi-K2.7-Code need?

Its published quantizations range from 283.04 GiB to 461.08 GiB in weights alone, before KV cache and runtime overhead.

Kimi-K2.7-Code on RTX PRO 6000 Blackwell Max-Q Workstation Edition

Can I run Kimi-K2.7-Code on a RTX PRO 6000 Blackwell Max-Q Workstation Edition?

Q: Can I run Kimi-K2.7-Code on a RTX PRO 6000 Blackwell Max-Q Workstation Edition?

No. No published quantization of Kimi-K2.7-Code fits a RTX PRO 6000 Blackwell Max-Q Workstation Edition at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of Kimi-K2.7-Code fits RTX PRO 6000 Blackwell Max-Q Workstation Edition at any context we compute, with q4_0 KV. The smallest shipped quantization is 283.04 GiB in weights alone, against 89.28 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— 96 GB card, 89.28 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◐
UD-IQ4_XS	461.08 GiB	462.2	462.3	462.5	462.8	463.4	464.6
UD-Q3_K_M	431.78 GiB	432.9	433.0	433.2	433.5	434.1	435.3
UD-IQ3_S	390.04 GiB	391.2	391.3	391.4	391.7	392.3	393.5
UD-IQ3_XXS	351.00 GiB	352.2	352.2	352.4	352.7	353.3	354.5
UD-IQ2_M	296.14 GiB	297.3	297.4	297.5	297.8	298.4	299.6
UD-IQ2_XXS	296.00 GiB	297.2	297.2	297.4	297.7	298.3	299.5
UD-IQ1_M	283.04 GiB	284.2	284.3	284.4	284.7	285.3	286.5

●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.

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.