DeepSeek-R1 Distill 1.5B · 11 GPUs measured first-party · llama.cpp Q4_K_M · Updated July 2026

What GPU Do You Need for DeepSeek-R1 Distill 1.5B?

DeepSeek-R1 Distill 1.5B compresses R1's reasoning style into the smallest possible package, and honestly, most of it gets lost in the compression. We measured it on 11 GPUs (llama.cpp, Q4_K_M): 635 tok/s on the RTX PRO 6000 Blackwell, ~2GB peak VRAM, and 149 tok/s even on a T4.

Benchmarked weights: bartowski/DeepSeek-R1-Distill-Qwen-1.5B-GGUF

Our honest take: skip it unless the job is tiny. At 1.5B, the reasoning traces this distill inherited from R1 are mostly ceremony: it thinks out loud, but the thoughts don't reliably improve the answer the way they do at 14B and up. Where it earns a slot is really light pipeline inference: cheap classification or extraction where you want a hint of chain-of-thought structure and you're processing at enormous volume. If you just need tiny-and-fast glue with direct answers, Qwen3 0.6B does that with fewer wasted tokens; if you want reasoning that matters, the 7B distill is the real starting point.

The numbers. 635 tok/s at the top and a ~2GB floor mean deployment cost is near zero: every card in our database fits it, and the H200's 5.05 tok/W is among the best efficiency figures we've logged. But remember what reasoning models do to effective speed: if a trace burns 500 thinking tokens before a 50-token answer, your perceived throughput divides by ten. At this size that overhead is the whole story. Which is exactly why we'd route real work to a bigger distill.

Our verdict

DeepSeek-R1 Distill 1.5B: 635 tok/s peak, ~2GB floor, runs on literally anything, but at this size the reasoning is mostly theater. Use it for very light pipeline inference at volume, and step to the 7B distill the moment answers start to matter.

FAQ

Is DeepSeek-R1 Distill 1.5B worth running?

Only for light, high-volume pipeline inference where cost dominates. The reasoning traces at 1.5B rarely change the answer, for anything user-facing, the 7B distill is the honest minimum in this family.

How much VRAM does it need?

~2GB measured peak at Q4_K_M: every GPU in our database clears it, including 4GB cards and old datacenter silicon like the T4 (149 tok/s).

Why is it slower in practice than the tok/s suggests?

It generates thinking tokens before answering, R1-style. A long trace before a short answer divides your effective throughput several-fold, factor that in when comparing it to direct-answer models of the same size.

DeepSeek 1.5B or Qwen3 0.6B for pipeline work?

Qwen3 0.6B if you want direct answers at maximum volume, it skips the reasoning overhead. The 1.5B distill only wins when a little chain-of-thought structure measurably helps your task.

What's the most efficient card for it?

The H200 measured 5.05 tok/W, but that's academic. This model can't stress anything. An L4-class card at 47W measured draw (189 tok/s) is the economical batch choice.