DeepSeek / Model selection guide
DeepSeek models:
versions, API names and downloads.
DeepSeek models include current hosted services, downloadable releases such as V4 and R1, smaller distilled models and specialist families. Start by choosing how you will run the model, then match its exact name to the task and license.
For DeepSeek’s own API, the current choices are deepseek-flash, serving DeepSeek-V4.1-Flash, and deepseek-v4-pro, serving DeepSeek-V4-Pro-0813. A downloadable checkpoint or a model on another host can have a different identifier and availability.
By China AI Chat Team · Core sources reviewed ; OCR-2 sources checked · Independent, source-based guide · Our research method
Which DeepSeek model should you choose?
Choose the deployment route first, then apply these selection criteria. They are guidance, not new test results.
| Your requirement | Start here | Check before committing |
|---|---|---|
| Chat without building an application | Official DeepSeek assistant | The modes and tools visible in your account. Do not infer its exact serving checkpoint from the API model table below. |
| Read images through the official API | Current Flash model | Supported image formats and vision instructions, then test the kind of image you actually use. |
| Build a text or coding application | Evaluate current Flash and Pro on the same small task set | Correctness, missed instructions, latency and total billed usage under matched thinking settings. |
| Run a model on your own hardware | An exact downloadable checkpoint, potentially a smaller distill | Runtime support, precision, weight storage, memory and context overhead. Parameter count alone does not establish that it fits. Check the release license and runtime requirements. |
| Reproduce an older result | The original repository revision or a host that explicitly offers it. See how model names differ. | An accepted legacy alias can serve a replacement. Preserve the original prompt, configuration and dependencies too. |
Define success before testing: valid structured data, a source-supported answer or a code change that passes your tests. Try representative and difficult cases; context size and advertised benchmark scores cannot guarantee success on your workload.
For a small example of how to inspect outcomes, our four-task DeepSeek–ChatGPT comparison publishes its prompts and first answers. It compares consumer apps, not the API models or downloadable weights listed here.
Current DeepSeek API models
This table describes DeepSeek’s own hosted API. It does not identify the model serving a particular consumer-chat response or list every downloadable release.
| API name | Model version | Input | Context / max output |
|---|---|---|---|
deepseek-flash | DeepSeek-V4.1-Flash | Text and images | 1M tokens / 384K tokens |
deepseek-v4-pro | DeepSeek-V4-Pro-0813 | Text | 1M tokens / 384K tokens |
Both current API models support thinking and non-thinking modes; thinking is the documented default. Flash accepts text and images, while the listed Pro model accepts text. Image input means a model can inspect an image; it does not mean this endpoint generates pictures. For an image payload, two first responses and token checks, use our tested DeepSeek Vision API guide.
Choosing between these two first-party API IDs? Our DeepSeek Flash vs Pro comparison shows six matched tasks, all first responses, the high-thinking code failure and rate-based costs. It is a narrow API test, not a ranking of every release.
The 1M context and 384K maximum output are provider-stated capacities, not recommended request sizes or a promise of perfect recall. Keep room for generated output and check the chosen mode’s parameters. Send a first DeepSeek API request for setup, examples and response checks.
References: current API model mapping and thinking-mode controls. Use our pricing guide and token-cost calculator for costs rather than assuming the larger model name determines your bill.
The main DeepSeek model families
Use this map to distinguish releases you may encounter in tutorials, model repositories and hosting services. “Earlier release” refers to its place in the family; it does not mean its download has disappeared. Current official API availability is shown separately above.
| Family or release | What it covers | Access and selection point |
|---|---|---|
| V4.1-Flash | Current Flash release with text and image input and text output. | Available through the official Flash API mapping and as published weights. Hosted settings and self-hosted configuration are separate. |
| V4-Pro-0813 | Current Pro checkpoint for text generation and reasoning. | Matches the current official Pro API mapping. Its dated repository differs from the earlier unsuffixed V4-Pro download. |
| Earlier V4 releases | Earlier V4-Pro, V4-Flash and Flash-0731 checkpoints. | Relevant when reproducing an existing deployment. An old Flash API alias no longer guarantees these weights are served. |
| V3 family | Earlier general text releases, including V3, V3-0324 and V3.2. | Use the exact release card for mode support and licensing. These names are not the current pair of first-party API identifiers. |
| R1 and R1-0528 | Reasoning-oriented releases with downloadable weights. | The full model and its distilled derivatives are different choices. A third-party R1 endpoint follows that host’s availability and settings. |
| R1 distilled models | Smaller models trained using R1-generated material. The original series includes Qwen-based 1.5B, 7B, 14B and 32B releases, and Llama-based 8B and 70B releases. | Start at the official release list. Check the base model, exact checkpoint, runtime and upstream license notices. |
The later R1-0528-Qwen3-8B is a separate Qwen3-based distillation, not one of the original six distills. A distilled model is not simply a compressed download of the full R1 model. Our Windows/Ollama walkthrough tests one Q4_K_M package derived from this later Qwen3-based distill, with the exact tag, first answers and CPU observations.
Specialist families solve different tasks
- DeepSeek-Coder-V2 is an earlier coding-focused family with Base and Instruct variants. Evaluate its exact release against your code task; “Coder” alone is not proof that it beats a newer general model.
- DeepSeek-VL2 covers visual understanding, including document and chart tasks.
- DeepSeek-OCR 2 is a separate downloadable document-reading checkpoint for image-to-text or Markdown extraction. Its official setup documents an NVIDIA CUDA route and renders PDF pages as images. DeepSeek’s published API model table does not list OCR-2 as a selectable model; neither that API nor the consumer PDF tool should be assumed to use these weights. Check the output against the page, especially tables and reading order. See our original DeepSeek-OCR 2 document-to-Markdown test for the first output and verification checklist.
- Janus-Pro combines multimodal understanding and image generation. It is a separate downloadable family, not evidence that DeepSeek’s current chat-completions API can generate images. See our Janus-Pro demo test and first image gallery for the practical route.
This is a task-oriented map, not a complete archive or performance ranking. Start with the verified DeepSeek model organization for the complete published catalog.
One model can appear under different kinds of name
Suppose a tutorial tells you to use Flash. Before copying its settings, identify which of these it means:
- API identifier:
deepseek-flashis the string accepted by DeepSeek’s current API. Its underlying version can change. - Downloadable repository:
deepseek-ai/DeepSeek-V4.1-Flashidentifies a published model repository. Record the repository revision when reproducing a local setup. - Third-party service identifier: another host chooses its own model ID and serving configuration. Use that provider’s endpoint, credentials, specification and price list.
A dated suffix helps distinguish releases, but an API alias is not a permanent version pin. Record the exact provider, identifier, date and settings when comparing results. Our DeepSeek overview explains the separate consumer app and official-access routes.
For the consumer workflow, see reading PDFs in DeepSeek Chat. Its upload tools and test results do not establish the capabilities of every downloadable checkpoint or API endpoint.
Before downloading: check the release and its license
Follow the official model-card link for the exact checkpoint. Read its usage instructions, weight license and any upstream notices; code and weights do not always share a license.
- Current V4.1-Flash and V4-Pro-0813: their linked cards state that both repository code and weights use the MIT license.
- Original V3 versus later V3 releases: the original separates MIT code from a DeepSeek Model License for weights. V3-0324 and V3.2 cards state MIT for code and weights. Do not transfer one release’s terms to another.
- R1 distills: the R1 repository’s license section identifies the Qwen- and Llama-derived models and their upstream terms. Read those notices alongside the release license.
- Specialists: Coder-V2, VL2 and Janus-Pro have separate code/weight-license documentation. For Janus-Pro, the card’s short metadata label and body differ; consult the linked weight license rather than relying only on a catalog badge.
Next, match the runtime and hardware requirements in that release’s instructions. Quantization reduces numerical precision and can change resource needs and output behavior; it is different from distillation. Check who produced a quantized file and which original revision it derives from.
This model catalog remains source-based and does not establish memory requirements across releases. The separate local R1 distill guide reports one measured Windows/CPU setup. Running a model yourself also leaves you responsible for the surrounding application, access controls and infrastructure costs.
What happened to older DeepSeek API names?
The current official reference accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp as legacy names, but routes them to V4.1-Flash. They do not preserve the old Flash or experimental vision model.
DeepSeek’s changelog documented retirement of deepseek-chat and deepseek-reasoner for July 24, 2026. Its later update says V4-Pro service continues beyond the previously announced September 14 shutdown. Use the current changelog together with the current API table when following an older tutorial.
These statements concern DeepSeek’s hosted service. They do not establish whether a downloaded checkpoint still works or whether another provider offers an older release. We did not probe retired endpoints for this guide.
Common questions about DeepSeek models
Is R1 the same as DeepThink?
No. R1 identifies a model family; DeepThink is a control in the consumer assistant. A button label alone does not identify the exact backend serving a response. Record the visible product and mode when the model version is not exposed.
Can I use a DeepSeek model for free?
Consumer-assistant access, model-download permission and inference costs are separate. The official assistant is advertised as free; hosted API calls are billed, and self-hosting requires computing resources. See DeepSeek access and pricing for the relevant route.
Which DeepSeek model is best?
There is no universal recommendation from this reference alone. Input type, deployment constraints, cost and success on your actual tasks should narrow the shortlist. Our six-task matched API comparison helps with the current hosted Flash and Pro IDs, but does not rank downloadable releases or prove an overall winner.
Where should I get the model files?
Start with DeepSeek’s verified model organization and the official repository linked from the release card. Check the exact revision, license and runtime instructions before choosing a third-party quantization or serving package.
Explore more DeepSeek guides
- DeepSeek AI overview — understand the product and find official access.
- DeepSeek R1 distill on Windows — set up one exact local package and inspect its first answers.
- DeepSeek API setup — send a first request and inspect its response.
- DeepSeek pricing and calculator — estimate costs from defined token counts.
Core source review: September 21, 2026; OCR-2 addition checked September 24, 2026. This guide organizes official specifications and release documentation; it does not present provider benchmarks as ChinaAI test results. Model aliases and service availability can change. How we research · Report a correction.