DeepSeek OCR 2 — Hardware Requirements & GPU Compatibility
Vision
DeepSeek OCR 2 is DeepSeek's compact vision-language model built for optical character recognition and document understanding, totaling 3.4 billion parameters with about 1.2 billion active per token through its mixture-of-experts design. Only the active parameters compute per token, keeping inference fast, though the full weight set still needs to fit in memory; at this size that's within reach of a consumer GPU or laptop once quantized. It uses a DeepEncoder V2 vision architecture that reasons over document layout semantically instead of scanning images in a fixed pattern. The model has an 8K token context window, suited to single-document OCR passes. It is released under the Apache 2.0 license, allowing unrestricted commercial and research use, and was published in January 2026 as DeepSeek's OCR successor.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 8K)Best on 24 GB (Q4_K_M · ~504 tok/s · 8K)Macs from 8 GB (Q4_K_M · ~69 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run DeepSeek OCR 2
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/sabafallah/DeepSeek-OCR-2-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 21.2 GB free · 8K context
Download: deepseek-ocr-2-Q4_K_M.gguf1.95 GBsabafallah
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 24 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
Runs on RTX 3060 12GB (9 GB headroom) · Q4_K_M
- Generation speed
- ~229 tok/s
- generation speed
- Cost per 1M output tokens
- $0.072
- per 1M output tokens
How Much VRAM Does DeepSeek OCR 2 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 1.9 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 8K context · ~600 tok/s | |
| 1.9 GB | — | 3-bit · i-quant medium — Importance-weighted 3-bit, medium | Fits: 8K context · ~585 tok/s | |
| 2.1 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 8K context · ~562 tok/s | |
| 2.3 GB | — | 4-bit · i-quant compact — Importance-weighted 4-bit, compact | Fits: 8K context · ~535 tok/s | |
| 2.3 GB | — | 4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear | Fits: 8K context · ~523 tok/s | |
| 2.5 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 8K context · ~504 tok/s | |
| 2.8 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 8K context · ~457 tok/s | |
| 3.2 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 8K context · ~418 tok/s | |
| 3.8 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 8K context · ~369 tok/s | |
| 7.2 GB | — | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 8K context · ~221 tok/s |
Selected Q4_K_M
4-bit medium quantization2.8GBat 8K context
- Weights
- 2.0 GB
- KV cache
- 0.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 21.2 GB free · ~504 tok/s
≥50 tok/s needs ≥69 GB/s · RTX 5090 ~712, RTX 3090 Ti ~504
Which Hardware Can Run DeepSeek OCR 2?
Q4_K_M · 2.5 GBDeepSeek OCR 2 (Q4_K_M) requires 2.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 4+ GB is recommended. Using the full 8K context window can add up to 0.4 GB, bringing total usage to 2.8 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run DeepSeek OCR 2, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).
≤ 11 GB
≤ 11 GB tier →fits · 5.2 GB free · up to 8K ctx
16 GPUs · 1 Mac class · 2 device families
RTX 3080~414Simulate output on RTX 3080RTX 2080 Ti~354Simulate output on RTX 2080 TiRTX 3070 Ti~351Simulate output on RTX 3070 TiM3*8 GB~69Simulate output on iMac 24" M3 (8 GB)Orin Nano~73Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)iPhone~49Simulate output on iPhone 1720 GB
fits · 17.2 GB free · up to 8K ctx
1 GPU
RX 7900 XT~238Simulate output on RX 7900 XT32–40 GB
fits · 29.2 GB free · up to 8K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~712Simulate output on RTX 5090M5 Max36 GB~196Simulate output on Mac Studio M5 Max (36 GB)M532 GB~96Simulate output on MacBook Air 13" M5 (32 GB)Orin~140Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~90Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to DeepSeek OCR 2
DeepSeek family · 5 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Unlimited OCR 2Baidu | — | 3.3 GB | ~517 | same size |
| Jina Ocr V1jinaai | — | 3.3 GB | ~509 | same size |
| Unlimited OCRBaidu | — | 3.3 GB | ~519 | same size |
| Granite 4.0 MicroIBM | — | 3.7 GB | ~261 | same size |
| Llama 3.2 3B InstructMeta | — | 4.1 GB | ~309 | same size |
| Tiny Aya GlobalCohere | — | 9.4 GB | ~89 | same size |
Specifications
- Publisher
- DeepSeek
- Family
- DeepSeek
- Parameters
- 3.4B
- Active parameters
- 1.2B per token
- Architecture
- DeepseekOCR2ForCausalLM
- Context Length
- 8,192 tokens
- Vocabulary Size
- 129,280
- Release Date
- 2026-01-27
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does DeepSeek OCR 2 need?
DeepSeek OCR 2 requires 2.5 GB of VRAM at Q4_K_M, or 7.2 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 3.4B × 4.8 bits ÷ 8 = 2 GB
KV Cache + Overhead ≈ 0.5 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 0.8 GB (at full 8K context)
VRAM usage by quantization
What's the best quantization for DeepSeek OCR 2?
For DeepSeek OCR 2, Q4_K_M (2.5 GB) offers the best balance of quality and VRAM usage. Q5_K_M (2.8 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 1.9 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run DeepSeek OCR 2 on a Mac?
Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run DeepSeek OCR 2. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 2.5 GB of usable unified memory (RAM minus macOS overhead).
Can I run DeepSeek OCR 2 locally?
Yes — DeepSeek OCR 2 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is DeepSeek OCR 2?
At Q4_K_M, DeepSeek OCR 2 can reach ~388 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~504 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 0.9 GB active ÷ (8000 × 0.65) = 0.1667 ms, plus 12 layers × 0.055 ms = 0.6600 ms, so 1000 ÷ 0.8267 ms = ~1210 tok/s
Estimated speed at Q4_K_M (2.5 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of DeepSeek OCR 2?
At Q4_K_M, the download is about 2.03 GB. The full-precision BF16 version is 6.78 GB. The smallest option (Q2_K) is 1.44 GB.
Which GPUs can run DeepSeek OCR 2?
52 consumer GPUs can run DeepSeek OCR 2 at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run DeepSeek OCR 2?
163 devices with unified memory can run DeepSeek OCR 2 at Q4_K_M (2.5 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.