By LLM Hardware team · Methodology ·
AI assembled the first DeepSeek table from DeepSeek's own model cards. VRAM figures follow the site formula, and speed figures are approximate and not all are cited yet.
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: DeepSeek-R1 14B at Q4 needs 10.4 GB. A 12 GB GPU fits it. An 8 GB GPU does not. R1 32B at Q4 needs 21.2 GB. R1 70B at Q4 needs 44 GB.
How we got it: 14 × 0.6 + 2 = 10.4. 32 × 0.6 + 2 = 21.2. 70 × 0.6 + 2 = 44. Ollama library on 11 October 2026 listed deepseek-r1 at 1.5b, 7b, 8b, 14b, 32b, 70b, and 671b.
Checked:
Key fact: DeepSeek-V4-Pro on Ollama is a cloud tag, not a local weight this page will put on a GPU.
How we got it: Ollama library tag deepseek-v4-pro:cloud on 11 October 2026. The command is ollama run deepseek-v4-pro:cloud.
Checked:
Key fact: DeepSeek-V4.1-Flash-1017 files sum to 510.3 GB. Q4 on the 763.2B total is 459.9 GB. Neither fits 128 GB or 512 GB after an OS reserve.
How we got it: Safetensors parameter total 763,205,315,794 and file byte sum 510,296,708,312 on deepseek-ai/DeepSeek-V4.1-Flash-1017, 17 October 2026. 763.2 × 0.6 + 2 = 459.9.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Updated 24 October 2026 · DeepSeek R1 distilled models · V4.1-Flash-1017 file size · Consumer GPU guide
DeepSeek R1 distills are the local files. R1 7B at Q4 is 6.2 GB and fits 8 GB. R1 14B at Q4 is 10.4 GB and needs 12 GB. R1 32B at Q4 is 21.2 GB and needs 24 GB. R1 70B at Q4 is 44 GB and needs 64 GB of unified memory. The verified commands are ollama run deepseek-r1:7b, deepseek-r1:14b, deepseek-r1:32b, and deepseek-r1:70b. ollama run deepseek-v4-pro:cloud is a cloud tag, not a file you download onto a GPU. In LM Studio, pick the Q4_K_M GGUF of the distill size that the formula says fits.
What is DeepSeek?
DeepSeek is both a Chinese AI company and a family of models. The models that matter for local inference are:
VRAM is calculated using the formula: params × bytes_per_param × 1.2 + 2. Q4 is params_B × 0.6 + 2. Q8 is params_B × 1.2 + 2. Add extra headroom for long reasoning chains — see the KV cache section below.
| Model | Size | Q4 VRAM | Q8 VRAM | Minimum Hardware | Best Hardware |
|---|---|---|---|---|---|
| R1-Distill-Qwen-1.5B | 1.5B | 2.9 GB | 3.8 GB | 4 GB GPU (Q4) | 8 GB GPU |
| R1-Distill-Qwen-7B | 7B | 6.2 GB | 10.4 GB | RTX 4060 8GB (Q4) | 12 GB GPU (Q8) |
| R1-Distill-Llama-8B | 8B | 6.8 GB | 11.6 GB | RTX 4060 8GB (Q4) | 12 GB GPU (Q8) |
| R1-Distill-Qwen-14B | 14B | 10.4 GB | 18.8 GB | Arc B580 12GB (Q4) | RTX 4090 24GB (Q8) |
| R1-Distill-Qwen-32B | 32B | 21.2 GB | 40.4 GB | RTX 4090 24GB (Q4) | 24 GB GPU (Q4 only) |
| R1-Distill-Llama-70B | 70B | 44 GB | 86 GB | Mac Studio 64GB | 128 GB unified |
| Full R1 671B (MoE) | 671B | 404.6 GB | — | Not a consumer GPU | Not practical |
R1 32B at Q4 is 21.2 GB, so a 24 GB GPU fits it and a 16 GB GPU does not. The RX 7900 XTX 24GB and RTX 3090 24GB also work. Use the VRAM Calculator for context-length-adjusted estimates.
Find your GPU or Mac below. Each card lists what distilled models fit and what does not.
RTX 4060 8GB
Runs:
Does not fit:
Q4 only for the 7B/8B — no headroom for Q8.
Intel Arc B580 12GB
Runs:
Does not fit:
Best budget pick for 14B at Q4. Arc ROCm support is improving but less mature than CUDA.
RTX 4060 Ti 16GB
Runs:
Does not fit:
16 GB fits the 14B distill at Q4 and does not fit it at Q8. Check current price on Amazon.
RTX 4070 12GB
Runs:
Does not fit:
12 GB fits the 14B distill at Q4 (10.4 GB) and does not fit Q8 (18.8 GB). Check current price on Amazon.
RTX 4070 Ti Super 16GB
Runs:
Does not fit:
Faster than a 4060 Ti, same 16 GB ceiling. The 14B distill at Q8 still does not fit.
RTX 4080 16GB
Runs:
Does not fit:
Bandwidth does not change the 18.8 GB miss. Check current price on Amazon.
RTX 3090 24GB (used)
Runs:
Does not fit:
24 GB fits the 32B distill at Q4 (21.2 GB) and does not fit Q8 (40.4 GB). Check current price on Amazon.
AMD RX 7900 XTX 24GB
Runs:
Does not fit:
Same 24GB ceiling as RTX 4090/3090. ROCm support works with Ollama and llama.cpp. See the AMD ROCm guide.
RTX 4090 24GB
Runs:
Does not fit:
Best single-GPU option at 24GB. Fastest 32B Q4 inference on a consumer card (~15 tok/s).
RTX 5090 32GB
Runs:
Does not fit:
32 GB fits the 32B distill at Q4 (21.2 GB) and does not fit it at Q8 (40.4 GB) or the 70B distill (44 GB). Check current price on Amazon.
Mac mini M4 16GB
Runs:
Does not fit:
Treat usable memory as about 12 GB. The 7B Q4 file fits. The 14B Q4 file at 10.4 GB is the ceiling and is tight.
Mac mini M4 24GB
Runs:
Does not fit:
A 24 GB Mac holds the 14B Q4 file. Q8 at 18.8 GB is tight once macOS is reserved. The 32B file does not fit.
Mac mini M4 Pro 48GB
Runs:
Does not fit:
After a 4 GB reserve a 48 GB Mac holds 32B at Q4. Q8 at 40.4 GB is the ceiling and is tight. The 70B file does not fit.
Mac Studio M4 Max 64GB
Runs:
Does not fit:
The minimum practical option for the 70B distill. ~600 GB/s bandwidth gives solid reasoning chain speed.
Mac Studio M4 Max 128GB
Runs:
Does not fit:
Runs the full distill lineup. 70B at Q8 fits with room to spare.
DeepSeek R1 models use chain-of-thought reasoning. Before outputting an answer, the model generates internal "thinking" tokens that work through the problem step by step. These are consumed by KV cache — the memory buffer that holds previously generated tokens so they do not need to be recomputed.
KV cache grows with every token
Each reasoning step adds to the KV cache. A complex coding or math problem can generate 500-2,000 thinking tokens before the answer begins. This temporarily adds 1-4 GB of KV cache on top of the base model weight memory.
Factor 2-4x more VRAM per query
The base VRAM figures in the table above are for model weights only. During active inference on complex prompts, peak VRAM usage can be 1.5-2x higher. Leave headroom — do not run a model that barely fits.
Responses are slower per query
Even at the same tokens-per-second speed, total response time is longer because the model generates more tokens. A 7B DeepSeek R1 response may take 3-4x longer than a Llama 3.1 7B response to the same prompt.
Context length multiplies this effect
Longer system prompts and conversation history also grow KV cache. For extended reasoning sessions, 16K-32K context, or agentic pipelines, consider stepping up one model size tier to ensure stability.
Practical rule: if the model weight fits in VRAM with less than 20% headroom, you will likely see out-of-memory errors or extreme slowdowns on complex reasoning tasks. See the VRAM Calculator to estimate KV cache at your context length.
Token generation speed is bottlenecked by memory bandwidth. Faster bandwidth means faster reasoning chains. The table below shows estimated Q4_K_M speeds. Note that for reasoning models, these speeds apply to the full token stream including thinking tokens — total wall-clock time per query is longer.
| Hardware | Bandwidth | 7B Q4 tok/s | 14B Q4 tok/s | 32B Q4 tok/s |
|---|---|---|---|---|
| RTX 5090 32GB | 1,792 GB/s | ~102 t/s | ~53 t/s | ~27 t/s |
| RTX 4090 24GB | 1,008 GB/s | ~58 t/s | ~30 t/s | ~15 t/s |
| RTX 4060 Ti 16GB | 288 GB/s | ~16 t/s | ~9 t/s | — |
| Mac Studio M4 Max 64GB | ~600 GB/s | ~34 t/s | ~18 t/s | ~9 t/s |
| Mac mini M4 Pro 48GB | ~273 GB/s | ~15 t/s | ~8 t/s | ~4 t/s |
Speed estimates based on: tokens/sec = bandwidth (GB/s) / model size in memory (GB). Real-world results vary by software version and system configuration. Mac mini M4 Pro bandwidth estimated at ~273 GB/s. Mac Studio M4 Max bandwidth estimated at ~600 GB/s.
Ollama
ollama run deepseek-r1:7b Easiest option. One command installs and runs the model. Use :1.5b, :7b, :8b, :14b, :32b, or :70b tags. GPU auto-detected on NVIDIA, AMD (ROCm), and Apple Silicon.
LM Studio
Search "deepseek-r1" in Discover GUI-based model browser and chat interface. Download GGUF quantizations directly from within the app. Best for non-technical users. Runs on Windows, Mac, and Linux.
Hugging Face + llama.cpp
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B Download GGUF files from the bartowski or unsloth community repos on Hugging Face. Run with llama.cpp for maximum control over quantization and context length.
For step-by-step installation instructions, see the how to run LLMs locally guide. For AMD GPU setup with ROCm, see the AMD ROCm guide.
RTX 4060 8GB
Runs the 7B and 8B distills at Q4. Solid reasoning capability for coding, writing, and math at a budget price. Upgrade if you want 14B.
Intel Arc B580 12GB
Best value path to the 14B distill. 12GB fits 14B at Q4. Arc ROCm and Ollama support has improved significantly — verify driver compatibility before buying.
RTX 4060 Ti 16GB
Best bang for DeepSeek use. 16GB runs the 14B at Q8, which is the highest quality reasoning you get at this price. Strongly recommended.
RTX 4090 24GB
Best single-GPU card for DeepSeek. Runs the 32B distill at Q4, which delivers near-full-R1 reasoning quality on complex tasks. Fast bandwidth (~1,008 GB/s) keeps thinking chains snappy.
Mac mini M4 24GB
Runs 14B at Q8 silently and efficiently. Unified memory avoids the VRAM vs RAM split. For 32B, step up to Mac mini M4 Pro 48GB. For 70B, Mac Studio M4 Max 64GB.
For a full cross-budget GPU comparison, see the best GPU for LLMs guide.
Best GPU for LLMs — Full Guide
All budget tiers from entry to workstation
LLM Quantization Explained
Q4 vs Q8 vs FP16 — when quality trade-offs matter
How to Run LLMs Locally
Step-by-step Ollama, LM Studio, llama.cpp setup
Apple Silicon for LLMs
M4, M4 Pro, M4 Max — which Mac for DeepSeek?
Ollama vs LM Studio
Which tool to use for running DeepSeek locally
CPU-Only Inference
Run DeepSeek-R1-Distill-8B without a GPU
AMD GPU ROCm Guide
Run DeepSeek on RX 7900 XTX with ROCm
Best LLMs to Run Locally
DeepSeek vs Qwen3, Phi-4, Llama 3.3 — 2026 picks
Qwen3 Hardware Requirements
Compare DeepSeek R1 Distill vs Qwen3 thinking
Llama 3.3 70B Hardware Guide
Hardware for the best open-weights 70B model
Qwen3 vs DeepSeek vs Llama 3.3
Decide which model family fits your hardware
How to Run DeepSeek R1 Locally
Step-by-step Ollama setup guide for all hardware tiers
DeepSeek V3 Hardware Requirements
Can you run DeepSeek V3 locally? 685B model needs 390 GB VRAM — and what to use instead.
VRAM Calculator
Calculate exact VRAM at your context length
R1 7B at Q4 is 6.2 GB and fits an RTX 4060 8 GB. R1 14B at Q4 is 10.4 GB and needs 12 GB. R1 32B at Q4 is 21.2 GB and needs 24 GB. R1 70B at Q4 is 44 GB and needs a 64 GB Mac. The full 671B model is 404.6 GB at Q4.
Yes for deepseek-r1:7b. That Q4 estimate is 6.2 GB. Q8 of the same 7B model is 10.4 GB and misses 8 GB. The 1.5B distill is 2.9 GB at Q4. The 14B distill is 10.4 GB and does not fit 8 GB.
DeepSeek R1 uses chain-of-thought reasoning, generating internal "thinking" tokens before giving its final answer. A single query can trigger hundreds to thousands of thinking tokens. Total token generation per query is 2-5x higher than a non-reasoning model of the same size, so each response takes noticeably longer even at the same tokens-per-second speed.
The full DeepSeek R1 is a 671-billion parameter Mixture-of-Experts model requiring data-center hardware. The distilled models (1.5B to 70B) are smaller dense models trained using knowledge distillation from the full R1, transferring its reasoning capability into a far smaller footprint. The distilled models are what most people run locally. They retain strong reasoning ability despite being orders of magnitude smaller.
No. ollama run deepseek-v4-pro:cloud is a cloud tag, not a local GGUF. DeepSeek-V4.1-Flash-1017 files summed to 510.3 GB on 17 October 2026, and Q4 on its 763.2B total is 459.9 GB. Neither fits 128 GB or 512 GB after an OS reserve. Do not size a gaming GPU for that checkpoint.
A 16 GB Mac fits the 7B Q4 distill (6.2 GB) after macOS takes a slice. The 14B Q4 file is 10.4 GB and is tight there. A 24 GB Mac holds 14B Q4. A 48 GB Mac holds 32B Q4 (21.2 GB). The 70B Q4 file is 44 GB and needs 64 GB once macOS is reserved.
Check VRAM requirements for DeepSeek models, or compare hardware options.
LLM Quantization Explained: Q4, Q8, F16
Run DeepSeek on less VRAM by choosing the right quantization level.
Running 70B Models Locally
Hardware requirements and realistic expectations for large models.
KV Cache and VRAM: How Context Length Blows Up GPU Memory
Why long context windows eat VRAM fast and how to manage it.
Ollama vs llama.cpp: Which Should You Use?
Comparing the two most popular local LLM runtimes.
VRAM figures on this page come from the site’s VRAM formula and model cards. The sitewide formula and the full source list are on the methodology page. For this guide we used:
Spot a number that does not match the linked source? Contact us and we will update the guide.