What GPU do I need for DeepSeek-R1 Distill 32B?

24GB minimum. Measured peak was ~21GB at Q4_K_M. RX 7900 XTX, RTX 3090/4090, or the 5090 (32GB) for extra context headroom.

What's the recommended two-model setup?

This model plus Qwen3 32B (or QwQ) on the same prompt, answers reconciled. Different labs, different strengths by field, the cross-check catches what each misses. One 24GB card runs them sequentially.

Do I need a datacenter GPU for it?

No. Dense 32B generation is bandwidth-bound: the B300 measured 83 tok/s, while the consumer RTX 5090 hit 71 tok/s on the sister dense 32B, 85% of the output. Datacenter rates only make sense for batch serving.

How bad is the reasoning-token overhead at this size?

Real but worth it: traces run hundreds to thousands of tokens, so at 83 tok/s a hard problem takes tens of seconds of thinking. At this size the trace demonstrably improves answers. This is the tier where you pay the tax and profit.

DeepSeek 32B or QwQ 32B?

Same job, same hardware class, near-identical measured speed (both ~83 tok/s peak). Ideal answer: both, as panel members, two different reasoning lineages disagreeing productively is the whole point of consensus.

DeepSeek-R1 Distill 32B · 10 GPUs measured first-party · llama.cpp Q4_K_M · Updated July 2026

What GPU Do You Need for DeepSeek-R1 Distill 32B?

DeepSeek-R1 Distill 32B is the flagship of the distill family, the closest you get to full R1-style reasoning in a package a 24GB card can hold. Measured on 10 GPUs (llama.cpp, Q4_K_M): 83 tok/s on the B300, ~21GB peak VRAM.

Benchmarked weights: bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF

82.9tok/s

Fastest: NVIDIA B300

measured, 3-run llama-bench

~21GB

VRAM needed (measured peak)

GPU-independent, applies to every card

GPUs measured

same pinned harness

0.36tok/W

Most efficient: NVIDIA H100 80GB HBM3

real power sampling, not TDP

What GPU Do You Need for DeepSeek-R1 Distill 32B?, tok/s, fastest 10

NVIDIA B300

82.9 tok/s

NVIDIA B200

78.55 tok/s

NVIDIA H100 80GB HBM3

75.8 tok/s

NVIDIA H200

75.8 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

66.06 tok/s

NVIDIA A100 40GB SXM4

43.55 tok/s

NVIDIA A100 80GB SXM4

43.11 tok/s

NVIDIA L40S

34.45 tok/s

NVIDIA A10G

23.5 tok/s

NVIDIA L4

11.78 tok/s

Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.

DeepSeek-R1 Distill 32B. Measured generation speed by GPU

NVIDIA B30082.9

NVIDIA B20078.55

NVIDIA H100 80GB HBM375.8

NVIDIA H20075.8

NVIDIA RTX PRO 6000 Blackwell Workstation Edition66.06

NVIDIA A100 40GB SXM443.55

NVIDIA A100 80GB SXM443.11

NVIDIA L40S34.45

NVIDIA A10G23.5

NVIDIA L411.78

GPU	tok/s	Prompt t/s	tok/W	Avg power
NVIDIA B300	82.9	1363	0.24	341.7 W
NVIDIA B200	78.55	2593.9	0.2	392.3 W
NVIDIA H100 80GB HBM3	75.8	2280.5	0.36	209.8 W
NVIDIA H200	75.8	2257.9	0.31	243.9 W
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	66.06	3428.6	0.25	260.1 W
NVIDIA A100 40GB SXM4	43.55	1166.1	0.27	160.5 W
NVIDIA A100 80GB SXM4	43.11	1181.3	0.23	185.6 W
NVIDIA L40S	34.45	2332.1	0.18	194.7 W
NVIDIA A10G	23.5	1032.5	0.16	146.9 W
NVIDIA L4	11.78	664.5	0.19	60.9 W

The pairing we actually recommend. Here's the play this model was born for: DeepSeek 32B plus a Qwen 32B-class model on the same prompt, two strong brains from different labs, each double-checking the other. That two-model setup produces higher-quality output than any single model either family offers, and one 24GB card hosts the whole rotation. It's the strongest version of our consensus argument: don't hunt for the one biggest model you can fit; assemble models that are better in different fields and make them audit each other.

Hardware value, measured. Dense 32B inference is bandwidth-bound and single-stream, and the chart shows what that means for spending: the B300 tops out at 83 tok/s while, on the sister dense Qwen3 32B, an RTX 5090 measured 71 tok/s, 85% of the datacenter flagship from a consumer card. Rent big iron for batch serving; for a personal reasoning panel, a 5090, or a 24GB card at half the speed, is the rational buy. Budget for the trace tax too: at 23 tok/s (A10G) a long think is a minute of waiting.

Our verdict

DeepSeek-R1 Distill 32B: 83 tok/s peak, ~21GB floor, and the best reason we know to own a big consumer card, pair it with a Qwen 32B in a two-model consensus and you get output quality neither achieves alone, all on hardware that sits under your desk.

FAQ

What GPU do I need for DeepSeek-R1 Distill 32B?: 24GB minimum. Measured peak was ~21GB at Q4_K_M. RX 7900 XTX, RTX 3090/4090, or the 5090 (32GB) for extra context headroom.
What's the recommended two-model setup?: This model plus Qwen3 32B (or QwQ) on the same prompt, answers reconciled. Different labs, different strengths by field, the cross-check catches what each misses. One 24GB card runs them sequentially.
Do I need a datacenter GPU for it?: No. Dense 32B generation is bandwidth-bound: the B300 measured 83 tok/s, while the consumer RTX 5090 hit 71 tok/s on the sister dense 32B, 85% of the output. Datacenter rates only make sense for batch serving.
How bad is the reasoning-token overhead at this size?: Real but worth it: traces run hundreds to thousands of tokens, so at 83 tok/s a hard problem takes tens of seconds of thinking. At this size the trace demonstrably improves answers. This is the tier where you pay the tax and profit.
DeepSeek 32B or QwQ 32B?: Same job, same hardware class, near-identical measured speed (both ~83 tok/s peak). Ideal answer: both, as panel members, two different reasoning lineages disagreeing productively is the whole point of consensus.