GPUFits
Models / DeepSeek-R1 671B
DeepSeek-R1 671B VRAM Requirements & GPU Pairing
DeepSeek-R1 is the totem of open reasoning models — and a black hole of hardware demand: a 671B-total, 37B-active MoE (256 routed experts + 1 shared, 8 active per token), MIT licensed. It uses MLA compressed-latent attention: KV = layers × (kvLoraRank + qkRopeHeadDim) × bytes, just ~1.2GB at 8K — the same league as an 8B model. KV is not the wall; weights are: Q4_K_M totals ~413.6GB, and no consumer single card comes close.
Only three realistic paths exist: unsloth dynamic quants (1.58-2.51 bit, ~161-250GB, needing multi-A100 or dual-M3-Ultra-class memory), the official distilled versions (R1-Distill-Qwen-32B runs on a 24GB card), or the API. With 37B active parameters, even when it fits, speed is single-to-double-digit tok/s (8×A100 theoretically ~44 tok/s). Full-blood R1 is a toy for enthusiasts and labs, not a rational purchase — but as the benchmark for what reasoning models can do, it is mandatory knowledge.
Architecture Specs
| Total parameters | 671B |
| Active parameters(MoE: VRAM follows total params, speed follows active params) | 37B |
| Layers | 61 |
| KV heads(MLA latent compression: KV = layers × (kvLoraRank + qkRopeHeadDim) × bytes — the standard GQA formula does not apply) | — |
| Head dim | — |
| Native context | 128K |
| License | mit |
| KV bytes per token (fp16) | 68.6 KB |
MLA architecture: KV is computed as 61 layers × (512+64)-dim compressed latents. The KV column above uses the MLA formula — not comparable to standard GQA models.
VRAM by Quant Tier (@8K context, fp16 KV)
| Quant | Weights | Total (incl. 1.5GB runtime overhead) |
|---|---|---|
| Q8_0 | 713.8 GB | 715.9 GB |
| Q6_K | 551.1 GB | 553.1 GB |
| Q4_K_M | 411.0 GB | 413.1 GB |
| Q3_K_M | 335.5 GB | 337.6 GB |
| Q2_K | 265.9 GB | 268.0 GB |
KV cache and runtime overhead do not change across quants; longer contexts grow the KV part linearly — adjust context length in the GPU checker tool.
GPU Verdict Matrix (Q4_K_M @8K)
| GPU | Usable VRAM | Verdict | Recommended quant | Theoretical speed | |
|---|---|---|---|---|---|
| RTX 3060 12GB | 12.0 GB | Not feasible | — | — | Try it → |
| RTX 3090 | 24.0 GB | Not feasible | — | — | Try it → |
| RTX 4070 Ti Super | 16.0 GB | Not feasible | — | — | Try it → |
| RTX 4090 | 24.0 GB | Not feasible | — | — | Try it → |
| RTX 5090 | 32.0 GB | Not feasible | — | — | Try it → |
| RTX A6000 | 48.0 GB | Not feasible | — | — | Try it → |
| A100 80GB | 80.0 GB | Not feasible | — | — | Try it → |
| H100 80GB | 80.0 GB | Not feasible | — | — | Try it → |
| RX 7900 XTX | 24.0 GB | Not feasible | — | — | Try it → |
| Mac mini M4 Pro (48GB) | 36.0 GB | Not feasible | — | — | Try it → |
| Mac Studio M4 Max (64GB) | 48.0 GB | Not feasible | — | — | Try it → |
| Mac Studio M3 Ultra (96GB) | 72.0 GB | Not feasible | — | — | Try it → |
Theoretical speed = bandwidth × 0.75 ÷ per-token weight bytes (active params for MoE); real-world results vary with framework/driver/CPU, ±30%.
FAQ
- What hardware does full DeepSeek-R1 actually need?
- Q4_K_M is ~413.6GB: at least 6×A100 80GB or 4×H100. The 1.58-bit dynamic quant (~161GB) still needs 2×H100 or ~3 M3 Ultra 96GB machines' memory. No consumer single/dual-card path exists.
- Why is DeepSeek-R1's KV cache so small?
- MLA latent attention: each layer stores a 512+64-dim compressed latent vector instead of full K/V — ~1.2GB at 8K, the same league as an 8B dense model. That is why a 671B model's KV does not explode.
- How can a regular user access R1-level ability?
- Two pragmatic paths: the official R1-Distill-Qwen-32B (tight on 24GB) inherits most of the reasoning style; or just use the API — far cheaper than buying hardware for full R1.
Related guides
- Running DeepSeek-R1 Locally: The Complete Hardware Guide
- Multi-GPU Setup for Local LLMs: NVLink, PCIe & Layer Splitting
- MoE Model Hardware Requirements: Total vs Active Parameters
Try DeepSeek-R1 671B in the GPU compatibility checker →
Data verified 2026-09-01