Can I run CodeLlama-34b-instruct-hf on a 12GB card?

No. No published quantization of CodeLlama-34b-instruct-hf fits a 12GB card at any context with f16 KV cache.

How much VRAM does CodeLlama-34b-instruct-hf need?

Its published quantizations range from 13.23 GiB to 33.39 GiB in weights alone, before KV cache and runtime overhead.

Does CodeLlama-34b-instruct-hf fit in 12GB of VRAM?

Not at these settings. No indexed quantization of CodeLlama-34b-instruct-hf fits 12GB card at any context we compute, with f16 KV. The smallest shipped quantization is 13.23 GiB in weights alone, against 11.16 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— 12 GB card, 11.16 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	33.39 GiB	35.0	35.8	37.3	40.3	46.3	58.3
Q6_K	25.78 GiB	27.4	28.2	29.7	32.7	38.7	50.7
Q5_K_M	22.20 GiB	23.8	24.6	26.1	29.1	35.1	47.1
Q5_0	21.64 GiB	23.3	24.0	25.5	28.5	34.5	46.5
Q5_K_S	21.64 GiB	23.3	24.0	25.5	28.5	34.5	46.5
Q4_K_M	18.83 GiB	20.5	21.2	22.7	25.7	31.7	43.7
Q4_K_S	17.83 GiB	19.5	20.2	21.7	24.7	30.7	42.7
Q4_0	17.74 GiB	19.4	20.1	21.6	24.6	30.6	42.6
Q3_K_L	16.55 GiB	18.2	18.9	20.4	23.4	29.4	41.4
Q3_K_M	15.17 GiB	16.8	17.6	19.1	22.1	28.1	40.1
Q3_K_S	13.60 GiB	15.2	16.0	17.5	20.5	26.5	38.5
Q2_K	13.23 GiB	14.9	15.6	17.1	20.1	26.1	38.1

●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 there are no speeds on this page

A capacity is not a card. Whether a model fits depends only on memory, so every figure above holds for any 12GB accelerator. How fast it runs depends on memory bandwidth, which varies several-fold between cards of the same capacity — so putting a tokens-per-second number here would be inventing one. Pick a specific card from hardware and the speed column appears.

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, the KV cache grows linearly with context and, past about 32K, becomes larger than the weights for many models.