Open weights · Apache 2.0
DeepSeek OCR-2
A roughly 3B MoE vision-language model (570M active) for document understanding, with weights under Apache 2.0.
Key Specifications
- Context window
- 8K tokens
- Released
- Parameters
- 3B, 0.57B active
- Inputs
- Text, image
Hosted API pricing
| Provider | Input /1M tokens | Output /1M tokens | Cost at 10M in + 2M out | |
|---|---|---|---|---|
Novita
|
$0.03 | $0.03 | $0.36 | View |
Heads up: Base-tier, on-demand rates per 1M tokens; cached, batch and long-context tiers excluded. A provider may serve a shorter context or a quantized build than the creator's release. Verify before provisioning. More on how we price.
Estimated cost to self-host
DeepSeek OCR-2 needs about 4 GB of GPU memory at 4-bit with its full 8K context. The cheapest rental that fits is 1× GTX 1660 Super at about $43 a month.
That costs the same as roughly 1B tokens a month on Novita's API. Self-hosting is more expensive below that volume.
| Precision | Memory | Cheapest fit (Full 8K context) | Cost /mo | Break-even vs API |
|---|---|---|---|---|
| 4-bitINT4 / FP4 |
4 GB
|
$43
|
1B tokens /mo
|
|
| 8-bitFP8 / INT8 |
6 GB
|
$43
|
1B tokens /mo
|
|
| 16-bitFP16 / BF16 |
9 GB
|
$43
|
1B tokens /mo
|
Estimates based on median on-demand rates for Nvidia GPUs. Memory is weights plus KV cache for one request, using FP8 KV cache where supported and FP16/BF16 otherwise. Break-even assumes a 5:1 input-to-output ratio. No guarantee of runtime support, usable performance, or that a matching quantized build exists. How we estimate costs.
Similarly priced models
The models nearest DeepSeek OCR-2 by blended rate, each at its own cheapest provider.
| Model | Blended /1M | Input /1M | Output /1M | Context | Cutoff | vs DeepSeek OCR-2 |
|---|---|---|---|---|---|---|
Llama 3.1 8B
Meta
|
$0.025 | $0.02 | $0.05 | 131K | Dec 2023 | −17% |
|
DeepSeek OCR-2
This model
DeepSeek
|
$0.03 | $0.03 | $0.03 | 8K | ||
Qwen3.7-Flash
Alibaba Cloud
|
$0.0467 | $0.03 | $0.13 | 1M | +56% | |
GPT-OSS-20B
OpenAI
|
$0.0483 | $0.03 | $0.14 | 131K | Jun 2024 | +61% |
Command R7B
Cohere
|
$0.0583 | $0.04 | $0.15 | 128K | +94% | |
Gemma 3 12B Instruct
Google Cloud
|
$0.0583 | $0.05 | $0.10 | 128K | Aug 2024 | +94% |
|
Llama 3.2 1B
Meta
|
$0.0583 | $0.03 | $0.20 | 128K | Dec 2023 | +94% |
|
GPT-OSS-120B
OpenAI
|
$0.0633 | $0.04 | $0.18 | 131K | Jun 2024 | +111% |
Nemotron 3 Nano 30B A3B
Nvidia
|
$0.075 | $0.05 | $0.20 | 262K | Jun 2025 | +150% |
|
Llama 3 8B
Meta
|
$0.0833 | $0.05 | $0.25 | 8K | Mar 2023 | +178% |
|
Qwen3-Coder-30B-A3B
Alibaba Cloud
|
$0.0917 | $0.06 | $0.25 | 262K | +206% |
Prices are USD per 1M tokens at each model's cheapest listed provider. Blended is the cost of 10M input plus 2M output tokens, spread over the 12M.
Frequently Asked Questions
What is DeepSeek OCR-2 good for?
Document parsing and OCR. Apache 2.0 weights, small enough for a single consumer GPU, at the lower-cost end of DeepSeek's lineup.
When is DeepSeek OCR-2 not a good fit?
General chat, coding or agent work: it has no tool calling, and each call takes only a short document.
What is the cheapest way to run DeepSeek OCR-2?
Hosted, unless you push serious volume. Novita charges $0.03 in / $0.03 out per 1M tokens. The cheapest rental that fits is 1x GTX 1660 Super at $43 a month, which costs the same as about 1B tokens a month on that API.
Can I self-host DeepSeek OCR-2?
Yes. The weights are Apache 2.0 licensed. At 4-bit it needs about 4 GB of GPU memory, which starts at roughly $43 a month on the cheapest rental that fits.
More from DeepSeek
| Model | Context | Input /1M | Output /1M |
|---|---|---|---|
|
DeepSeek V4 Flash
Open weights
|
1M | $0.08 | $0.25 |
|
DeepSeek V3.2
Open weights
|
131K | $0.25 | $0.80 |
|
DeepSeek V3.1
Open weights
|
131K | $0.27 | $1.00 |
|
DeepSeek V3.1 Terminus
Open weights
|
131K | $0.27 | $1.00 |
|
DeepSeek V3
Open weights
|
128K | $0.27 | $1.12 |
|
DeepSeek V4.1 Flash
Open weights
|
1M | $0.30 | $1.20 |
|
DeepSeek V4 Pro
Open weights
|
1M | $0.45 | $0.89 |
|
DeepSeek V3 Turbo
Open weights
|
64K | $0.40 | $1.30 |