Last fact-checked: August 23, 2026. DeepSeek changes quickly, so this page separates current API facts, dated release history, company reporting, and our own clearly labelled inferences. It is a fact-check, not a substitute for the live official documentation.
Current snapshot: DeepSeek’s documented hosted API catalog contains three explicit model IDs: deepseek-v4-flash, deepseek-v4-pro, and the experimental multimodal deepseek-v4-flash-vision-exp. Flash and Pro are text-only. Vision Exp accepts text plus images for visual understanding; it is not an image generator. The image-only Files API can upload an image once and reuse its file_id, but it is not general PDF, DOCX or ZIP storage.
Disclosure: Chat-Deep.ai is an independent resource. It is not DeepSeek, DeepSeek.com, the official DeepSeek app, or the official developer platform.
How we verify DeepSeek facts
- Official DeepSeek sources: API documentation, release notes, model cards, policies, terms and the transparency center.
- Official public authorities: regulators and government security directions.
- Reliable reporting: used for company ownership, funding and IPO reporting when no public issuer filing exists.
- Our inference: used only when it is explicitly labelled and supported by cited facts.
We do not treat a model’s answer about itself, an anonymous social-media account, a third-party host, a copied logo or a crypto listing as proof. Time-sensitive model names, prices, limits and policies should be rechecked at the official source before a production decision.
DeepSeek facts at a glance
| Claim | Verdict | Why |
|---|---|---|
| DeepSeek is a Chinese AI company founded by Liang Wenfeng | True | Official legal policies identify the operator; reliable reporting documents the founder and 2023 origin. |
| DeepSeek is one model | False | DeepSeek has multiple model families, checkpoints and hosted products. |
| The current hosted API has only Flash and Pro | False | The current catalog also includes experimental deepseek-v4-flash-vision-exp. |
| All three V4 API models accept images | False | Only Vision Exp accepts image input; Flash and Pro remain text-only. |
| DeepSeek now has a general document Files API | False | The current Files API is for supported images; it does not document PDF, DOCX or ZIP input. |
| Vision Exp generates images | False | It is documented for visual understanding, not hosted image generation. |
| Hosted DeepSeek and a local checkpoint have the same privacy profile | False | Hosted processing follows DeepSeek’s service policies; local processing depends on the whole deployment. |
| DeepSeek is completely open source in every sense | Incomplete | Official open weights exist, but that is not the same as every hosted component, dataset or derivative being open. |
| DeepSeek guarantees accurate answers | False | DeepSeek’s own terms warn that outputs can be inaccurate and require human review in high-impact uses. |
| DeepSeek has a public stock ticker | False | It remains a private company on the sources reviewed. |
Company and ownership facts
Fact 1: DeepSeek is operated by a Hangzhou-based Chinese company
DeepSeek’s privacy policy and terms identify Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the provider or controller, with a registered address in China. Reliable reporting identifies Liang Wenfeng as founder and describes the company’s origin in 2023.
Fact 2: High-Flyer was the original financial base, but old sole-owner wording is unsafe
Liang also co-founded the quantitative hedge fund High-Flyer, which historically backed DeepSeek. Later reporting described outside financing in 2026. Without a current public cap table, say that DeepSeek was founded and historically financed through High-Flyer and that outside funding was later reported; do not claim permanent sole ownership.
Fact 3: The company, consumer chatbot, API and open checkpoints are different things
“DeepSeek” may mean the company, a model family, the official web/app service, the developer API, or an open-weight checkpoint. A third-party website or local runtime can use a DeepSeek model without being operated or endorsed by DeepSeek. Identify both the model and the provider.
Current model and API facts
Fact 4: The hosted V4 catalog now has three explicit IDs
| Model ID | Input | Best fit | Important boundary |
|---|---|---|---|
deepseek-v4-flash |
Text | Fast, cost-sensitive text and agent workloads | Rejects image content. |
deepseek-v4-pro |
Text | Harder reasoning, coding and agent workloads | Rejects image content; higher price and lower concurrency than Flash. |
deepseek-v4-flash-vision-exp |
Text + images | Experimental visual understanding and multimodal agents | No FIM; not an image generator; images add billed input tokens. |
All three are documented with a 1M context length, up to 384K output, thinking and non-thinking modes, JSON output, tool calls, Chat Completions, Responses and Anthropic-compatible access. Capability summaries can change. Test the exact endpoint and model combination you use.
Fact 5: Vision Exp adds image understanding without changing Flash or Pro
DeepSeek released deepseek-v4-flash-vision-exp on August 21, 2026 as an experimental multimodal model. DeepSeek says it retains Flash’s text capability and adds visual understanding, with a major gain for multimodal agent tasks. This does not make deepseek-v4-flash or deepseek-v4-pro multimodal; sending image content to those text models should be treated as an invalid request.
Supported image formats are JPEG, PNG, GIF and WebP. An application can send a public image URL, a Base64/data URL, a reusable Files API file_id, or supported inline file_data. A production gateway should validate model, message role, MIME type, source, byte size, image count and privacy policy before forwarding.
Current documented Vision limits
- Image URL: at most 8,192 characters.
- External URL or Base64 image: at most 32 MiB per image.
- Files API image referenced by
file_id: at most 64 MiB. - Inline request body: at most 48 MiB.
- Up to 600 images in a request.
- Total image bytes: up to 64 MiB without
file_id, and up to 200 MiB includingfile_id. - Maximum side: 8,192 px, reduced to 4,096 px when a request contains 15 or more images.
These are provider limits, not recommended defaults. Use fewer, smaller, purpose-specific images. Recheck the official Vision guide before implementation.
Fact 6: Files API means reusable image uploads, not general document support
The official Files API lets an application upload a supported image once and reuse the resulting file_id in Vision requests. DeepSeek describes the upload service as free to use, while the subsequent model request remains billable. It is not a Batch API and does not document general PDF, DOCX, spreadsheet or ZIP ingestion.
If source material is a document, extract approved text or render selected pages to validated images outside DeepSeek, then apply your privacy, copyright and accessibility review. Do not silently rename a PDF as an image.
Fact 7: Image understanding is not image generation
Vision Exp can inspect screenshots, photographs, charts and diagrams. The official hosted API sources reviewed here do not document an image-generation endpoint. DeepSeek has published other multimodal research and open models, including Janus, but that does not turn the hosted V4 Vision endpoint into an image generator.
Fact 8: Context and output maxima are capacities, not quality promises
The current V4 table documents 1M context and a maximum output of 384K. That does not guarantee perfect recall, reasoning, or economical use at those limits. Image inputs are converted into input tokens, with a documented maximum of 384 input tokens per image after processing. Control image count, resolution, text context and output length, then use returned usage fields for billing analysis.
Fact 9: Compatibility does not mean identical behavior
DeepSeek documents OpenAI- and Anthropic-compatible interfaces and a native Responses API. Compatibility reduces migration work; it does not guarantee identical parameter support, message shapes, tool behavior, errors or image serialization. Test the exact route. In particular, FIM is not available for Vision Exp, and general document blocks remain outside the documented image feature.
Fact 10: Legacy aliases are history, not current model identities
deepseek-chat and deepseek-reasoner moved between underlying model versions over time. DeepSeek announced their retirement after temporary V4 Flash mappings. Preserve them only in dated migration history. New integrations should use an explicit current ID and should not depend on unlisted alias fallback.
Pricing and concurrency facts
Fact 11: Vision Exp shares Flash’s token price class but images can increase total cost
The current pricing matrix lists the same token rates for Flash and Vision Exp. Equal rates do not guarantee an equal request bill: each Vision image contributes input tokens. Pro has a higher token price class. Prices can change, so use the official pricing page for a live budget.
Fact 12: Weekend pricing follows the Beijing calendar
Effective August 23, 2026, Saturday and Sunday are entirely off-peak, determined by the current date in Beijing—not by the caller’s local calendar. On weekdays, DeepSeek’s peak windows are 01:00–04:00 UTC and 06:00–10:00 UTC; other hours are off-peak. Convert the request timestamp to the provider’s published schedule before estimating cost, and never hard-code a permanent weekend assumption without a documented effective date.
Fact 13: Concurrency is an account-level ceiling, not a throughput target
The current limit page lists 2,500 concurrent requests for Flash, 500 for Pro and 2,500 for Vision Exp, counted at account level across keys. Start far below the ceiling, protect downstream systems, and tune from real latency, 429s, queue age and cost.
Open-weight and privacy facts
Fact 14: Official open weights exist, but “everything is open source” is too broad
DeepSeek publishes official weights and technical material for several model families. Check the exact repository, checkpoint and license. A weight license does not automatically describe the hosted app/API, training data, every derivative, or third-party service.
Fact 15: Truly local inference changes data flow, but the surrounding stack still matters
A checkpoint running entirely on infrastructure you control does not need to send prompts to DeepSeek’s hosted API. The runtime, plugins, telemetry, logging, vector store and connected tools may still transmit data elsewhere. “Local UI” is not proof of local-only inference.
Fact 16: Hosted-service privacy is different
DeepSeek’s privacy policy describes collection that can include prompts, uploaded files, photos, chat history and device/network data depending on the service. It says the service is not designed or intended for sensitive personal data and describes processing and storage in the People’s Republic of China. Read the current policy and your organization’s requirements before sending confidential material.
Fact 17: Training opt-out is not a zero-retention promise
The policy describes an option concerning model improvement or training. That does not by itself mean zero collection, zero logs or immediate deletion. Treat training use, service processing and retention as separate questions.
Fact 18: Model output is not guaranteed fact or professional advice
DeepSeek’s terms warn that output can contain errors and should not be treated as professional advice. Thinking mode, long context, tools or vision can improve a workflow but cannot remove the need for source verification and qualified human review, especially for decisions affecting people’s rights or safety.
Government, business and identity facts
Fact 19: Restrictions exist, but “DeepSeek is banned everywhere” is false
Public authorities have taken different actions involving hosted services and government systems. The applicable rule depends on jurisdiction, organization, product path and date. A government-device restriction is not automatically a ban on every private user or every locally hosted checkpoint.
Fact 20: DeepSeek remains private and a similarly named token is not equity
DeepSeek has no public retail stock ticker on the sources reviewed. Reported private valuations or IPO preparation are not a completed listing. DeepSeek has also said it did not issue a cryptocurrency; a copied name or logo does not make a token company equity.
DeepSeek release timeline
| Date | Release or event | Why it matters |
|---|---|---|
| 2023 | DeepSeek founded | Liang Wenfeng established the Hangzhou AI company after building High-Flyer. |
| May 7, 2024 | DeepSeek-V2 | Modern efficiency story with Mixture-of-Experts, MLA and 128K context. |
| September 5, 2024 | DeepSeek-V2.5 | Combined the general chat and Coder-V2 lines. |
| December 26, 2024 | DeepSeek-V3 | Expanded the main general-model line. |
| January 20, 2025 | DeepSeek-R1 | Major reasoning-focused release; deepseek-reasoner exposed R1 at that time. |
| May 28, 2025 | DeepSeek-R1-0528 | Updated reasoning, JSON and function-calling behavior. |
| August 21, 2025 | DeepSeek-V3.1 | Unified thinking/non-thinking modes and strengthened agent/tool use. |
| September 29, 2025 | DeepSeek-V3.2-Exp | Experimental long-context attention direction preceding V3.2. |
| December 1, 2025 | DeepSeek-V3.2 | Last main V3 release before V4. |
| April 24, 2026 | DeepSeek-V4 Preview | V4 Pro and Flash, 1M context, explicit V4 IDs and open weights. |
| July 24, 2026 | Announced legacy-alias retirement deadline | New code should use explicit V4 IDs rather than deepseek-chat or deepseek-reasoner. |
| July 31, 2026 | V4 Flash 0731 API update | Flash API public-beta update and native Responses support. |
| August 13, 2026 | V4 Pro 0813 GA | Pro rollout across App, Web and API with native Responses support. |
| August 21, 2026 | V4 Flash Vision Exp + Files API | Third hosted model ID, native image understanding, reusable image uploads and three supported API families. |
| August 21, 2026 | Harness 0.1.1 announcement | DeepSeek announced out-of-box Vision support in Harness; exact public package status must be checked separately. |
| August 23, 2026 | Weekend pricing rule | Every Saturday and Sunday by Beijing date became off-peak; weekday peak windows remain time-based. |
Frequently asked questions
What are the current DeepSeek API model IDs?
The current documented IDs are deepseek-v4-flash, deepseek-v4-pro, and experimental deepseek-v4-flash-vision-exp. Recheck the official model and pricing pages before deployment because the catalog can change.
Does DeepSeek support image input?
Yes, through deepseek-v4-flash-vision-exp. Flash and Pro remain text-only. Vision Exp accepts supported images by public URL, Base64/data URL, Files API file_id, or supported inline file_data.
Can DeepSeek generate images through the V4 API?
No hosted V4 image-generation endpoint is documented in the official sources reviewed here. Vision Exp understands images; it does not generate them.
Does the Files API support PDF or DOCX documents?
No. The current Files API is documented for supported image uploads and reuse in Vision requests. Do not treat it as general document storage, a document-analysis API or a Batch API.
Does Vision Exp cost the same as Flash?
The official pricing matrix lists the same token-rate class for Vision Exp and Flash. A Vision request can still cost more because every image contributes billed input tokens.
When is DeepSeek API usage off-peak?
As of August 23, 2026, every Saturday and Sunday by Beijing date is off-peak. On weekdays, peak windows are 01:00–04:00 and 06:00–10:00 UTC; other hours are off-peak. Verify the official schedule before budgeting.
Is DeepSeek open source?
“Open weight” is the safer artifact-specific description. Official checkpoints and licenses exist, but the hosted service, all training assets and every derivative are not automatically covered by one blanket claim.
Is Chat-Deep.ai the official DeepSeek website?
No. Chat-Deep.ai is an independent guide and is not affiliated with, endorsed by or operated by DeepSeek.
Official sources for the current API facts
- V4 Flash Vision Exp and Files API release
- Vision guide and image limits
- Files API
- Models and pricing
- Rate limits and isolation
- Responses API
- API change log
For implementation details, continue to the DeepSeek API guide, the current model catalog, the Vision Exp guide, and the pricing page.