DeepSeek-R1 Distill Qwen 14B VRAM Requirements

Weights are only half the question

A 14.8B model at Q4_K_M is 7.7 GB of weights. That part is fixed — it is the file on disk, and it does not change while the model runs. What changes is the key-value cache, which grows with every token of context you give it.

DeepSeek-R1 Distill Qwen 14B spends 192.0 KB per token. At 8K that is 1.5 GB, comfortably smaller than the weights. At its full 128K it is 24.0 GB — larger than the weights themselves. That is the trap: the model loads fine and then fills the card once a conversation gets long.

The reason the cache is as small as it is comes down to grouped-query attention. DeepSeek-R1 Distill Qwen 14B has 40 query heads but only 8 key-value heads, and the cache is sized by the latter. Calculators that use the query head count overstate it 5-fold, which is where the wildly pessimistic numbers people quote usually come from. Despite the name this is a Qwen2 architecture — R1's reasoning distilled into Qwen2.5 14B, not a DeepSeek model shape. That matters for planning: reasoning models spend thousands of tokens thinking before they answer, so the KV cache, not the weights, is what usually decides whether a long session fits. Its config carries a sliding_window value but sets use_sliding_window to false, so none is applied here.