DeepSeek Models: Every Model, Spec, Price and Release Date

DeepSeek is an AI company that publishes its large language models both as a paid API and as open weights on Hugging Face, and its API line-up was renamed and consolidated several times in 2026. This hub lists every DeepSeek model in DeepSeek's public API documentation and Hugging Face model cards as of 15 November 2026, with API IDs, release dates, context windows, prices, licences, availability and retirement dates. It is written for developers, buyers, architects and anyone who needs to know which DeepSeek model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from DeepSeek's own documentation or its official Hugging Face cards, cited in the sources list.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses DeepSeek's public API documentation as of November 2026. Model details change, so confirm against DeepSeek's current documentation.

Key takeaways

• Two API models are current. DeepSeek's pricing page lists deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813); older API names are retired or routed to deepseek-flash.
• Both current models list 1M context. The pricing page gives a 1M-token context length and up to 384K output tokens, with thinking mode on by default.
• Prices change by time of day. Off-peak rates are half the peak rates; deepseek-flash costs $0.30 peak or $0.15 off-peak per million cache-miss input tokens.
• Weights are published under MIT. The V4.1-Flash-1017, V4-Pro-0813 and R1 cards state the MIT License; the original DeepSeek-V3 weights use a custom Model Agreement.
• Legacy names are being phased out. deepseek-chat and deepseek-reasoner were scheduled for discontinuation on 24 July 2026, and V4-Flash was retired on 10 September 2026.

What is DeepSeek?

DeepSeek develops the DeepSeek V-series and R-series large language models. It offers them through the DeepSeek API, which accepts both OpenAI-format requests at https://api.deepseek.com and Anthropic-format requests at https://api.deepseek.com/anthropic, and through its own app and web chat. DeepSeek also publishes model weights on its Hugging Face organisation, deepseek-ai. API names have changed over time: from 2024 to mid-2026 two stable names, deepseek-chat and deepseek-reasoner, were silently upgraded to each new model, and from April 2026 DeepSeek moved to versioned names such as deepseek-v4-pro and then to deepseek-flash, which always points to the latest Flash model.

The main table lists the models in the 2026 V4 generation plus the earlier models named in this page's scope. Facts are as of 15 November 2026.

Current DeepSeek models at a glance

As of 2026-11-15. Source: DeepSeek pricing page, change log and Hugging Face cards. Models are listed by release date, newest first. Category is the DeepSeek model line.

The pricing page states thinking and non-thinking modes for both current models, JSON output, tool calls, the Responses API, the Anthropic API format and chat prefix completion (beta). FIM completion (beta) works in non-thinking mode only. Vision input is listed for deepseek-flash and not for deepseek-v4-pro. See Open-Weight vs Closed Models Explained.

Key facts for the main DeepSeek models

As of 2026-11-15. Identical fields for every model; "Not stated by DeepSeek" means not stated by DeepSeek on the pages checked (2026-11-15). Parameter counts are those on the Hugging Face model card or size box.

The V4.1-Flash-1017 card recommends temperature 1.0, top_p 1.0 and max_tokens of at least 256K, and the V4-Pro-0813 card recommends up to 384K output for high and max reasoning. The V4-Pro-0813 weights include a DSpark speculative decoding module. Hugging Face file sizes can exceed the stated parameter counts: the V3 card explains its 684B listing as the 671B main model plus 14B of multi-token prediction weights. See Mixture of Experts Explained.

Features stated across current models

• Thinking: non-thinking and thinking modes, thinking by default; effort levels low, high and max were added on 13 August 2026.
• API formats: OpenAI Chat Completions, the Responses API (native from 31 July and 13 August 2026) and the Anthropic API format.
• Context caching: cache-hit input is priced separately from cache-miss input.
• Application APIs: web search, file input, URL reading and model safety for deepseek-flash from 13 November 2026, with search, fetch, OCR, speech to text and text to speech also offered as standalone APIs.
• Concurrency: the pricing page lists a limit of 2,500 for deepseek-flash and 500 for deepseek-v4-pro.

DeepSeek model families and naming

• V4.1 Flash: the current Flash line, with native image input. The API name deepseek-flash always points to the latest Flash version; DeepSeek-V4.1-Flash-1017 has been served since 17 October 2026.
• V4 Pro: the larger 1.7T-parameter text model, released in preview on 24 April 2026 and made generally available as DeepSeek-V4-Pro-0813.
• V4 Flash: released on 24 April 2026, given an official API release in public beta on 31 July 2026 and retired on 10 September 2026; V4-Flash-Vision-Exp was an experimental image-capable variant.
• V3 generation: V3 (December 2024), V3-0324, V3.1 (hybrid reasoning, August 2025), V3.1-Terminus, V3.2-Exp and V3.2, each served under deepseek-chat and deepseek-reasoner.
• R1: the reasoning line, introduced on 20 January 2025 and upgraded to R1-0528 on 28 May 2025. DeepSeek also published six R1-Distill models built on Qwen2.5 and Llama 3 base models.
• Version suffixes: a four-digit suffix such as 0813 or 1017 is the month and day of the update. See LLM Naming and Versioning Explained.

When was each DeepSeek model released?

As of 2026-11-15. Dates are from the DeepSeek API change log, in date order. Earlier V2-era entries are omitted.

The latest model update is DeepSeek-V4.1-Flash-1017, on 17 October 2026. The latest Pro release is DeepSeek-V4-Pro-0813, made generally available on 13 August 2026.

How much do DeepSeek models cost?

As of 2026-11-15. Price per million tokens, shown with a "$" sign on the pricing page (the page does not name the currency code). DeepSeek does not list a batch rate or separate cache-write price; the Batch and Cache write columns are kept for parity with other provider hubs. Source: DeepSeek pricing page.

• Peak hours: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours, weekends and Chinese public holidays are off-peak, at half the peak rate. Peak and off-peak pricing took effect at 16:00 UTC on 16 August 2026.
• Input price: "Input" is the cache-miss price and "Cached input" the cache-hit price.
• Legacy names: requests to deepseek-v4-flash and deepseek-v4-flash-vision-exp are routed to V4.1-Flash-1017 and billed at the Flash price.
• Tool charges: search (text or image) and web fetch cost $0.75 per 1,000 calls each; file parsing and OCR $0.225 per 1,000 pages; text to speech and speech to text are priced per character and per hour, with a limited-time 50% discount shown.
• Billing: fees are taken from the topped-up or granted balance, granted balance first, and DeepSeek states it may adjust prices.

See Model Pricing Explained and How to set spend limits on LLM API keys.

Where are DeepSeek models available?

• First-party: the DeepSeek API (OpenAI and Anthropic request formats) and DeepSeek's app and web chat.
• Open weights: the deepseek-ai organisation on Hugging Face hosts the V4.1-Flash, V4-Pro, V4-Flash, R1 and V3 weights, plus V4 base models and speculative decoding (DSpark) variants.
• Cloud platforms: partner cloud availability is not stated by DeepSeek on the pages checked (2026-11-15).
• Regional processing: not stated by DeepSeek on the API pages checked; the privacy policy states that personal data is collected, processed and stored in the People's Republic of China.

Because the weights are published, organisations can self-host DeepSeek models instead of calling the API. Self-hosting moves operation, patching and access control to the organisation. See the Self-Hosted LLM Security Guide.

Restricted-access DeepSeek models

As of 2026-11-15. Source: DeepSeek change log and pricing page.

No DeepSeek model is described as restricted to verified or invited organisations on the pages checked.

What is the context window of DeepSeek models?

Both current API models, deepseek-flash and deepseek-v4-pro, list a 1M-token context length and up to 384K output tokens on the pricing page. The V4.1-Flash-1017 card gives 1M tokens as the recommended context window. The earlier open-weight models DeepSeek-V3 and DeepSeek-R1 list 128K. DeepSeek does not tier prices by prompt length. See Context Windows Explained.

Is any DeepSeek model open source?

DeepSeek publishes weights for its main models. The DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro-0813 and DeepSeek-R1 cards state the MIT License, and the R1 card states that it permits commercial use, modification and derivative works, including distillation. The DeepSeek-V3 card states MIT for the code and a custom DeepSeek Model Agreement for the weights, which permits commercial use. The R1-Distill models are covered by their base models' licences: Apache 2.0 for the four Qwen2.5-based models and the Llama 3.1 or Llama 3.3 licence for the two Llama-based models. Training data is not published. See the Open-Weight Model Security Guide.

Model cards and system cards

As of 2026-11-15.

The V4.1-Flash-1017 card has no safety section; it gives a contact address for raising issues. See Model Cards and System Cards Explained.

Benchmarks

As of 2026-11-15. Provider-reported and independent results are kept in separate columns. Scores are as DeepSeek reports them on the Hugging Face cards and change log.

The V4-Pro-0813 card marks DSBench-FullStack and DSBench-Hard as internal DeepSeek test sets. Benchmark versions differ between cards, so scores are not directly comparable across models. See How to Read LLM Benchmarks.

Deprecations and retirements

As of 2026-11-15. Source: DeepSeek API change log. DeepSeek does not publish a fixed notice period on the pages checked.

The routing of retired V4-Flash names is described as temporary, so code that still sends deepseek-v4-flash should move to deepseek-flash. See the AI Model Retirement Guide.

Data handling and security

• Storage location: DeepSeek's privacy policy (last updated 22 December 2025) states that it collects, processes and stores personal data in the People's Republic of China.
• Training use: the privacy policy lists training and improving DeepSeek's models as a purpose of processing; it does not describe an opt-out.
• Retention: the policy states that personal data is kept for as long as the user has an account, with longer retention for legal and policy reasons.
• API customers: the policy states that data from end users of applications built on DeepSeek's open platform is not covered by it, and that the developer is responsible for its own privacy notice. Separate API retention or zero data retention terms are not stated by DeepSeek on the pages checked (2026-11-15).
• Model safety API: a model safety application API was released for deepseek-flash on 13 November 2026.
• Key handling: generic controls apply to every provider's API: keep keys out of code, scope and rotate them, and set spend limits.

See AI Data Retention and Deletion and the Foundation Model Selection Guide.

How to choose a DeepSeek model: factual criteria

• Modalities: deepseek-flash accepts image input; deepseek-v4-pro is text only.
• Price per token: compare cache-miss, cache-hit and output rates, and whether the workload can run in off-peak hours at half price.
• Concurrency: the stated limit is 2,500 for deepseek-flash and 500 for deepseek-v4-pro.
• Version pinning: deepseek-flash moves to each new Flash update; record the model version returned if results must be reproducible.
• Hosting model: API use means data is processed under DeepSeek's terms; self-hosting the MIT-licensed weights keeps data in the organisation's own environment.
• Hardware for self-hosting: V4-Pro has 1.7T and V4.1-Flash 763B total parameters, which sets the GPU memory needed.
• Licence: MIT for V4.1-Flash-1017, V4-Pro-0813 and R1; the Model Agreement for the original V3 weights; base-model licences for R1-Distill.
• Policy constraints: check internal and public sector rules on the use of the provider before adoption.

Practitioner checklist

• Replace any remaining deepseek-chat, deepseek-reasoner or deepseek-v4-flash calls with deepseek-flash or deepseek-v4-pro.
• Record the model version, such as DeepSeek-V4.1-Flash-1017, in the AI inventory, not only the API name.
• Review the DeepSeek change log on a schedule, since API names are re-pointed without a published notice period.
• Schedule batch-style workloads into off-peak hours where cost matters.
• Assess data residency and training-use terms against internal policy before sending regulated data to the API.
• For self-hosted weights, check the licence file in the repository and track model provenance.
• Store DeepSeek API keys in a secrets manager and set spend limits. See the LLMjacking Guide.

Standards and references

• DeepSeek: Models and pricing (accessed 2026-11-15)
• DeepSeek: API change log (accessed 2026-11-15)
• DeepSeek: DeepSeek-V4.1-Flash-1017 model card (accessed 2026-11-15)
• DeepSeek: DeepSeek-V4-Pro-0813 model card (accessed 2026-11-15)
• DeepSeek: DeepSeek-R1 model card (accessed 2026-11-15)
• DeepSeek: DeepSeek-V3 model card (accessed 2026-11-15)
• DeepSeek: deepseek-ai organisation on Hugging Face (accessed 2026-11-15)
• DeepSeek: Privacy policy (accessed 2026-11-15)

Related NHI Mgmt Group resources: Reasoning Models Explained, Multimodal Models Explained, Open-Weight Model Security Guide, LLM Security Guide.