Can the GeForce RTX 5070 run DeepSeek R1 (671B)?
Does not fit at Q4_K_M.
DeepSeek R1 (671B) is 671B parameters (MoE; all experts resident). At Q4_K_M the weights are about 375.0 GiB; with the 1.5 GiB runtime reserve that is 376.5 GiB against the card's 12GB, about 364.5 GiB more than the card has.
All three quants
12GB card| Quant | Weights | With reserve | Verdict | Headroom |
|---|---|---|---|---|
| Q4_K_Mthe everyday quant | 375.0 GiB | 376.5 GiB | Does not fit | short 364.5 GiB |
| Q8_0near-lossless | 662.4 GiB | 663.9 GiB | Does not fit | short 651.9 GiB |
| FP16full weights | 1249.8 GiB | 1251.3 GiB | Does not fit | short 1239.3 GiB |
Verdict rules: fits comfortably when weights plus reserve sit within 85% of VRAM; tight when they fit with little left for context; CPU offload when they exceed VRAM by up to 2x (runs, slowly, with layers in system RAM); no beyond that. The KV cache per token depends on the model's architecture and is not modelled here.
Not this card. These fit it comfortably at Q4.
Nothing in the consumer catalog fits DeepSeek R1 (671B) at Q4 on a single card; this is multi-GPU or datacenter territory, or a job for hosted inference (see PicksByModel for per-token pricing).
What the GeForce RTX 5070 does run well → · DeepSeek R1 (671B) on every card →
About the model
DeepSeek R1 (671B) by DeepSeek: 131,072-token native context, text, released 2025 under the MIT. Model card → · Hosted pricing on PicksByModel →