DeepSeek V3 / R1 VRAM requirements

How much VRAM does DeepSeek V3 / R1 need?

DeepSeek V3 / R1 needs 425 GB at Q4_K_M, 746 GB at Q8_0 and 1,404 GB at FP16 with 8,192 tokens of context and one request. Each total below is the weights plus the FP16 KV cache and the runtime overhead (0.5 GB plus 10%), and each row opens the calculator with that setting. What the GGUF names mean.

Fine-tuning DeepSeek V3 / R1? DeepSeek V3 / R1 VRAM for LoRA, QLoRA and full training.

DeepSeek V3 / R1 GGUF files on Hugging Face

The 5 GGUF files llama.cpp picks for DeepSeek V3 / R1 with -hf <repo>:<QUANT>, with the bytes the Hugging Face file list reported on 2026-09-29. Each total is that file plus the FP16 KV cache and 0.5 GB + 10%, one request. "vs estimate" compares the file with the bits-per-weight size the table above uses.

GB here is GiB (1024³ bytes), as everywhere on this site; Hugging Face shows the Q4_K_M file as 404.4 GB, in 10⁹ bytes.

Which DeepSeek V3 / R1 GGUF fits a 16, 24, 32, 48 or 80 GB GPU or a Mac

From the file sizes above: the largest file that fits each machine with an 8K context and 0.5 GB left free, the longest context it then has room for, and the longest context for Q4_K_M (377 GB). One request, FP16 KV cache. The Q4_K_M file is 9.04 GB under its bits-per-weight estimate; the estimate tables and the llama-server command on this page count the estimate, so they show less context.

Run DeepSeek V3 / R1 with llama.cpp

llama-server -hf unsloth/DeepSeek-V3-GGUF:Q4_K_M -c 163840

DeepSeek-V3-Q4_K_M/DeepSeek-V3-Q4_K_M-00001-of-00009.gguf, 404.4 GB, from unsloth/DeepSeek-V3-GGUF (checked 2026-09-29). -c 163840 is the longest Q4_K_M context on 8× H100 SXM (640 GB), layers split over the cards.

DeepSeek V3 / R1 at 8K, 32K, 128K and 160K (full) tokens of context

All 61 layers use multi-head latent attention. Each extra token of context adds 69 KB of FP16 cache per request. The last column is the smallest setup here, one card or a group, that holds it at Q4_K_M. How the KV cache works.

How fast DeepSeek V3 / R1 writes on a multi-GPU group

Tokens per second for one request with 8,192 tokens of context, estimated from memory bandwidth and the parameters read per token. No single card or machine holds it at Q4_K_M or FP8 (the precisions below), so these rows are the GPU groups here that do, in tensor parallel with their combined bandwidth. A dash means it does not fit at that precision. Try other settings in the speed calculator.

Longest context on a multi-GPU group

No single card or machine holds it at Q4_K_M, Q8_0 or FP8 (the precisions below), so these rows use GPU groups: how many tokens of context DeepSeek V3 / R1 fits on each, in one tensor-parallel group with one request, an FP16 KV cache and 0.5 GB left free. "Full" means the model's whole context window fits.

Model details

Parameters

684.5B (684,531,386,000)

Experts

256 routed experts, all loaded

Active per token

37B

Layers

All 61 layers use multi-head latent attention

Attention cache

576 values per token (compressed latent)

Context length

163,840 tokens

Published weights

641 GB (FP8)

On Hugging Face

deepseek-ai/DeepSeek-V3

Reported benchmarks

GPQA Diamond 59.1 · MMLU-Pro 75.9 · LiveCodeBench (Pass@1) 37.6 (chat model). Vendor-reported: from the maker’s own model card, not rerun here; read 2026-09-30.

DeepSeek V3 / R1 next to similar-sized models

The three models here closest to it in parameter count, at Q4_K_M with 8,192 tokens of context.