DeepSeek Models: Which One Should You Use?
For a new app or coding tool using DeepSeek’s API, start with V4.1-Flash. It accepts text and images and costs less than V4 Pro. If you want to run a model on your own computer, look at the smaller R1-derived models first; the full-size releases need substantially more hardware.
Just want to ask questions or work with a document? You can use DeepSeek’s official chat without downloading model files. The app and website offer their own interface; the API model names below are for developers choosing what powers an application.
V4.1-Flash vs V4 Pro: the current API choices
As of October 4, 2026, DeepSeek’s official model list and changelog document these two API options:
| Model and API name | Input | When to choose it |
|---|---|---|
DeepSeek-V4.1-Flashdeepseek-flash | Text and images | The starting choice for a new integration, including coding tools and questions about screenshots or charts. |
DeepSeek-V4-Pro-0813deepseek-v4-pro | Text | Keep or choose it when comparison on your own tasks shows a useful advantage that justifies the higher cost. |
Both support thinking and non-thinking modes, tool calls and a one-million-token context window. That gives an application room to supply substantial reference material, but does not guarantee that every detail will be used correctly.
Pro is not an automatic upgrade for difficult work. In its V4.1-Flash announcement, DeepSeek reports benchmark results ahead of V4 Pro.
If Pro already works well in your application, compare the two on examples that matter before switching. For a coding assistant, use the same bug or feature request and check whether the resulting code works. For document questions, check the answer against the source. Pay attention to the time and cost needed to get an acceptable result.
For the rates and a worked cost example, see DeepSeek API pricing. To connect your chosen model, follow the API setup guide.
R1 and smaller models for local use
DeepSeek-R1 is a reasoning model family with downloadable model files, also called weights. The full R1 and its later R1-0528 release are large deployments. For a first local setup, the smaller models trained using R1’s reasoning outputs are more practical candidates.
The original R1 model card lists distilled models in 1.5B, 7B, 8B, 14B, 32B and 70B sizes, based on Qwen or Llama. For a first local trial, consider the later R1-0528-Qwen3-8B if your computer has enough memory for an 8B model. These are separate smaller models, so an “8B R1” download will not produce the same results as full R1.
Choose a size your computer can run comfortably before chasing a higher score. The file format, compression and length of your conversation also affect memory use. Our DeepSeek system requirements guide explains how to match a download to your machine.
V4.1-Flash and V4 Pro also have published weights, but “Flash” does not mean a small laptop download. Use the official model card to check deployment support. A third-party service may host R1 or another older model for you; its model selection, pricing and data handling are separate from DeepSeek’s own API.
Before downloading, check the license for that exact release, including any terms inherited from its base model. There is no single license that covers every model carrying the DeepSeek name.
Other DeepSeek models worth knowing
The specialist families below are useful when you need their particular task or a model you can deploy yourself. They are not extra model choices in the current official API.
- Coder and Coder V2: earlier coding-focused releases. Coder V2 has full-size and Lite variants. Consider them for a specific self-hosted coding project; you do not need a model named “Coder” to get coding help from today’s general models.
- OCR and OCR 2: document image-to-text models. These are candidates for a pipeline that extracts text from scanned pages, rather than a general chat assistant.
- VL and VL2: models for understanding images, including questions about charts and documents.
- Janus and Janus-Pro: a separate family for image understanding and generation. Flash’s ability to read an image does not mean it can generate one.
- Math and Prover: mathematical reasoning and formal-proof research. Prover targets proofs in Lean; ordinary maths questions do not require setting up a theorem-proving model.
Where V3 and older model names fit
V3, V3.1, V3.2 and V2 are earlier generations. Their downloadable releases can still matter for an existing application or research comparison, even when DeepSeek’s hosted API has moved on. For the sequence of releases, see the DeepSeek model history.
Older tutorials also use deepseek-chat and deepseek-reasoner. These names changed the models they referred to over time; they do not identify a fixed V3 or R1 download. For a new integration, use a current API name from the table above. For a local model, use the exact repository and version listed in DeepSeek’s official Hugging Face account.