DeepSeek-R1 distill VRAM requirements: 1.5B to 70B

Short answer: DeepSeek’s six R1 distill checkpoints span very different memory tiers. The official repositories display approximately 3.56GB for Qwen-1.5B, 15.2GB for Qwen-7B, 16.1GB for Llama-8B, 29.5GB for Qwen-14B, 65.5GB for Qwen-32B and 141GB for Llama-70B in their full-precision repositories. Quantized community artifacts can be much smaller, but DeepSeek does not publish one official GGUF size table for this family.

DeepSeek-R1 distill models at a glance

Official distill repository	Base model published by DeepSeek	Official repository display	Compact-quant planning lane
DeepSeek-R1-Distill-Qwen-1.5B	Qwen2.5-Math-1.5B	3.56GB	4GB+; exact artifact and context decide
DeepSeek-R1-Distill-Qwen-7B	Qwen2.5-Math-7B	15.2GB	8GB–12GB starting lane
DeepSeek-R1-Distill-Llama-8B	Llama 3.1 8B	16.1GB	8GB–12GB starting lane
DeepSeek-R1-Distill-Qwen-14B	Qwen2.5-14B	29.5GB	12GB–16GB starting lane
DeepSeek-R1-Distill-Qwen-32B	Qwen2.5-32B	65.5GB	24GB–48GB lane
DeepSeek-R1-Distill-Llama-70B	Llama 3.3 70B Instruct	141GB	48GB+ lane

Raw 4-bit floors are not complete VRAM requirements

Using only the model labels, raw 4-bit arithmetic is approximately 0.75GB, 3.5GB, 4GB, 7GB, 16GB and 35GB for the six sizes. Real quantized artifacts can be larger because of mixed tensors, metadata and alignment. Resident memory must also include KV cache, runtime buffers, allocator overhead, the display workload and any concurrent sequences.

Distill size	Nominal 4-bit raw floor	Why the hardware tier is higher
1.5B	0.75GB	Runtime, cache and operating-system allocations dominate the tiny raw floor
7B	3.5GB	Long context and complete runtime state can exceed a 4GB device
8B	4GB	A 4GB label leaves no room before overhead
14B	7GB	8GB is tight; 12GB or 16GB provides more working room
32B	16GB	16GB is only the raw floor; 24GB is the first common compact-4-bit target
70B	35GB	48GB is the first common one-device compact-4-bit target

Which DeepSeek distill size should you choose?

• 1.5B: choose for the smallest hardware footprint and fastest experimentation; do not assume it matches larger-model reasoning quality.
• 7B or 8B: practical mainstream-GPU starting points. They use different Qwen and Llama bases, so parameter count alone does not make them interchangeable.
• 14B: a middle lane for 12GB/16GB hardware with a compact artifact and measured context.
• 32B: the first high-VRAM desktop lane, commonly beginning around 24GB for compact 4-bit use.
• 70B: a 48GB+ capacity problem before long context, larger quantizations or concurrency.

Context and reasoning output can dominate the margin

DeepSeek’s distilled models can generate long reasoning traces, so maximum output length, stop behavior, prompt template and generation settings affect real memory and latency. The 32B and 70B exact-model guides show how a 128K allocation can add tens of GiB of illustrative FP16 KV payload. Smaller variants have different layer and KV-head counts; do not reuse one model’s cache number for another.

Do not confuse distilled dense models with DeepSeek-R1 MoE

DeepSeek publishes the original DeepSeek-R1 as 671B total parameters with 37B activated and a 128K context. The distill models are dense Qwen/Llama checkpoints fine-tuned with R1-generated samples. Active parameters describe routing and compute in the MoE model; they do not turn the original model into a 37B stored-weight artifact.

A reproducible selection checklist

1. Write down the complete repository name—not just “DeepSeek R1.”
2. Record the exact artifact, revision and quantization.
3. Verify the intended tokenizer, prompt template and generation settings.
4. Set context, cache precision, batch, concurrency and maximum output.
5. Measure peak device and host memory on representative reasoning prompts.
6. Evaluate answer quality for the selected size and quantization.
7. Record prompt processing, generation speed and time to first token separately.

Official sources

• DeepSeek-R1 repository and official distill-model table
• Official Qwen-1.5B repository
• Official Qwen-7B repository
• Official Llama-8B repository
• Official Qwen-14B repository
• Official Qwen-32B repository
• Official Llama-70B repository
• HighVRAM methodology

Reviewed October 25, 2026. Repository totals are the official Hugging Face displays for DeepSeek’s full-precision checkpoints. Raw 4-bit values are planning arithmetic, not official quantized artifacts or measured peak memory. HighVRAM has not independently benchmarked every configuration. Capacity does not establish speed, reasoning quality or value.