Can RTX 5070 12GB run DeepSeek-R1?

NVIDIA · 12 GB GDDR7 · 672 GB/s bandwidth · DeepSeek · 684.5B (active 37B) · MoE · model ctx up to 128K

No — DeepSeek-R1 at Q4 needs ≈ 381.18 GB even at 4K context, beyond the RTX 5070 12GB's 12 GB GDDR7. It would only run with heavy CPU/disk offload. FP16 needs ≈ 1377.4 GB — does not fit on this card. Estimates come from memory-bandwidth math; rows tagged measured override estimates. Relative ranking is more reliable than absolute tok/s. Methodology.

Fit & speed by quant and context

Quant	Context	VRAM needed	Fits?	tok/s (decode)
Q4	4K	381.18 GB	No	—
Q4	8K default	384.88 GB	No	—
Q4	32K	407.08 GB	No	—
FP16	4K	1373.7 GB	No	—
FP16	8K default	1377.4 GB	No	—
FP16	32K	1399.6 GB	No	—

Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s means the model is far beyond this card (offload-only territory).

VRAM breakdown at 8K context

Quant	Weights	KV cache	Overhead	Total needed	RTX 5070 12GB VRAM
Q4	376.48 GB	7.4 GB	1 GB	384.88 GB	12 GB
FP16	1369.0 GB	7.4 GB	1 GB	1377.4 GB	12 GB

Other GPUs that run DeepSeek-R1

No catalog GPU fully fits DeepSeek-R1 at Q4/8K — see the model page for offload estimates.

All GPUs for DeepSeek-R1 →

Other models for the RTX 5070 12GB

• Qwen3-0.6B — ≈534.5 tok/s Q4 est.
• gemma-3-1b-it — ≈427.6 tok/s Q4 est.
• Llama-3.2-1B-Instruct — ≈356.4 tok/s Q4 est.
• Qwen3-1.7B — ≈213.8 tok/s Q4 est.
• Llama-3.2-3B-Instruct — ≈133.6 tok/s Q4 est.
• Mistral-7B-Instruct-v0.3 — ≈120.5 tok/s Q4 measured
• Phi-4-mini-instruct — ≈112.5 tok/s Q4 est.
• Phi-4-mini-reasoning — ≈112.5 tok/s Q4 est.

All models on the RTX 5070 12GB →