DeepSeek-OCR
- In Input modalities: TextImage
- Open weights
Prices
1 offering · prices as listedvanchin/deepseek-ocr| Conditions/ 1M tokens | Input | Output |
|---|---|---|
| Standard | ¥0.216 | ¥0.216 |
Prices from 阿里云百炼：模型价格 (opens in a new tab) (checked Nov 2, 2026)
Standard: the price a request with no conditions pays (the headline price elsewhere on the site).
Details
- Licence
- MIT
- Model ID
Not stated on DeepSeek’s pages: output modalities, file formats, output formats, max file size (mb), context, max output, structured output, languages.
Run locally
Weights: deepseek-ai/DeepSeek-OCR- Parameters 3.3B
- Active 1.1B (estimated) · 64 experts, 6 per token
- Native precision BF16
- Layers 12
- KV cache per token 60 KB
- Config positions 8K (config.json max_position_embeddings)
- Needs
- ≈ 7.2 GB (weights 6.2 (maker’s file) + KV cache 0.5 + overhead 0.5)
- Usable
- 21.6 GB (90% of its memory)
- Generation speed
- 181–271 tok/s (reads the active 1.1B per token)
- Longest context
- 8K (the model’s limit)
- First token
- about 0.04–0.06 s for a 2K prompt
| Precision | Weights | Needs at 8K | On your hardware | Longest context |
|---|---|---|---|---|
| BF16 (native)maker’s file | 6.2 GB | ≈ 7.2 | ●Runs well 181–271 tok/s | 8K |
| Q8_0 | 3.3 GB | ≈ 4.3 | ●Runs well 294–441 tok/s | 8K |
| FP8 | 3.1 GB | ≈ 4.1 | ●Runs well 307–460 tok/s | 8K |
| Q6_K | 2.6 GB | ≈ 3.5 | ●Runs well 350–525 tok/s | 8K |
| Q5_K_M | 2.2 GB | ≈ 3.2 | ●Runs well 383–575 tok/s | 8K |
| Q5_0 | 2.1 GB | ≈ 3.1 | ●Runs well 391–587 tok/s | 8K |
| Q4_K_M | 1.9 GB | ≈ 2.9 | ●Runs well 421–632 tok/s | 8K |
| Q4_0 | 1.7 GB | ≈ 2.7 | ●Runs well 440–660 tok/s | 8K |
| IQ4_XS | 1.7 GB | ≈ 2.6 | ●Runs well 454–681 tok/s | 8K |
| Q3_K_Mquality drops | 1.5 GB | ≈ 2.5 | ●Runs well 475–713 tok/s | 8K |
| IQ3_XXSquality drops | 1.2 GB | ≈ 2.2 | ●Runs well 536–804 tok/s | 8K |
| Q2_Kquality drops | 1.2 GB | ≈ 2.1 | ●Runs well 541–812 tok/s | 8K |
| IQ2_XXSquality drops | 0.8 GB | ≈ 1.8 | ●Runs well 632–948 tok/s | 8K |
Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.
Compared with other hardware
Each at the highest precision it runs. Click one for details.
Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 3080; unified memory Apple M5 · 16 GB.
Run commands
vllm serve deepseek-ai/DeepSeek-OCR --max-model-len 8192
Memory to fine-tune
Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.
Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated
Cited scores
From third-party leaderboards, not measured by usNo cited leaderboard covers OCR models yet. We will cite one when its license allows.
Changes
All changes →- New OCR model
Sources
- deepseek-ai/DeepSeek-OCR model card (opens in a new tab) huggingface.co · checked Nov 3, 2026
- 阿里云百炼：模型价格in Chinese (opens in a new tab) help.aliyun.com · checked Nov 2, 2026
- GitHub - deepseek-ai/DeepSeek-OCR: Contexts Optical Compression · GitHub (opens in a new tab) github.com · checked Nov 5, 2026
- deepseek-ai/DeepSeek-OCR model card (opens in a new tab) huggingface.co · checked Nov 3, 2026
- 模型上下架与更新-大模型服务平台百炼-阿里云_大模型服务平台百炼(Model Studio)-大模型服务平台百炼(Model Studio)-阿里云帮助中心in Chinese (opens in a new tab) help.aliyun.com · checked Nov 4, 2026
- deepseek-ai/DeepSeek-OCR config.json (opens in a new tab) huggingface.co · checked Nov 1, 2026
- deepseek-ai/DeepSeek-OCR files (opens in a new tab) huggingface.co · checked Nov 4, 2026
DeepSeek’s official policies: data use, privacy, terms →
DeepSeek · China · deepseek.com