dalao.ai

DeepSeek

A large-language-model company with MIT-licensed open weights and a pay-as-you-go API priced by peak and off-peak hours.Reference profile

About DeepSeek

DeepSeek is a company that builds large language models. It offers a web chat, a mobile app, an API platform and the desktop agent app DeepSeek Harness. Model weights are published on Hugging Face under MIT, the DeepInfer inference framework is Apache-2.0 and Harness is MIT; the web chat, app and API service are closed source. The main API model name, deepseek-flash, is served by DeepSeek-V4.1-Flash-1017 with a 1M-token context; peak output is $1.2 per million tokens, half that off-peak (DeepSeek pricing page, read 2026-10-26). Latest model update: V4.1-Flash-1017, October 17, 2026.

EditorialUpdated 10 min read

What it is

DeepSeek is a company that builds large language models. As of 2026-10-26 its home page lists four entry points: a web chat, a mobile app, an API platform and a desktop agent app, DeepSeek Harness. A banner says V4.1-Flash has been updated to V4.1-Flash-1017 and that the API has added two endpoints, web search and web fetch.

Models are the core product. The API currently offers two model names: deepseek-flash is served by DeepSeek-V4.1-Flash-1017, and the pricing page maps deepseek-v4-pro to DeepSeek-V4-Pro-0813. Both have a 1M-token context and up to 384K output tokens per call. The September 10, 2026 release note calls V4.1-Flash the smallest model in a new architecture family: a 552B-parameter mixture-of-experts model with 8B active parameters for input and 16B for output, plus native image understanding.

Open and closed parts need to be read separately. Model weights are published under MIT on Hugging Face and ModelScope, and the January 2025 R1 release note explicitly allows distillation and commercial use. DeepInfer, the inference framework made public in October 2026, is Apache-2.0, and the code of the DeepSeek Harness desktop app is MIT. The web chat, app and API service are closed source, and neither the training data nor the server systems are public.

As of 2026-10-26 the deepseek-ai GitHub organization has 45 public repositories and 110,869 followers; the Hugging Face organization page lists 106 models and 151,513 followers. Inference: these numbers show how widely its models and code circulate among developers, and many third-party inference services also deploy its open weights, so the same model can be called through several channels, with price and availability varying by channel.

Stronger benchmark performance, faster task completion, and shorter outputs than the previous version.

DeepSeek-V4.1-Flash-1017 release note

Core capabilities

Long context and low prices are the two points most often compared. deepseek-flash has a 1M-token context and a 384K-token output cap, and supports JSON output, tool calls, the Responses API and an Anthropic-format API; chat prefix completion and FIM completion are in beta, and FIM works only in non-thinking mode. Without a parameter the model runs in thinking mode, and the docs explain how to switch to non-thinking mode.

API compatibility makes it easy to adopt. The first-call docs state that the API format is compatible with OpenAI and Anthropic: the OpenAI-format base_url is https://api.deepseek.com and the Anthropic-format one is https://api.deepseek.com/anthropic. With the base_url, key and model name changed, existing SDKs or tools that speak either API can call it. The docs' agent integrations section has setup guides for DeepSeek Harness, Claude Code, Codex, OpenCode, WorkBuddy/CodeBuddy and others.

Cache-hit pricing matters a lot for agents. The September 10 release note says that, compared with the previous generation, the V4.1-Flash KV cache needs only a quarter of the HBM and an eighth of the SSD storage, and notes that cache-hit charges often make up a large share of agent costs. On the pricing page, cache-hit input for deepseek-flash costs one fiftieth of cache-miss input.

V4.1-Flash-1017, updated on October 17, focuses on building complete projects from scratch. The release note says it uses reinforcement learning with self-feedback to improve end-to-end project development, scores higher on benchmarks, finishes tasks faster and produces shorter outputs, and is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy; Harness users are advised to upgrade to v0.2.1-rc.1 or later.

On the same day the web search (Search) and web fetch (Fetch) APIs opened to all users, powered by DeepSeek's own search engine, at $0.75 per 1,000 successful requests per endpoint. Both are billed separately from model calls, so an agent can search, fetch a full page and hand the result to the model without a third-party search service.

Who it's for / Not for

A fit for: developers and teams that call models heavily, especially for coding agents and long-document analysis, with long inputs and high cache-hit rates. The 1M context plus very low cache-hit prices suit tasks that reread the same code or documents: when the same prefix enters the model a second time, it is billed at the cache-hit rate. Teams that want to deploy open weights on their own machines, or need a license that allows commercial use and distillation, are also well served by the MIT weights.

A fit for: people already using Claude Code, Codex or OpenCode, who can switch to DeepSeek by changing only the base_url and model name, without changing tools.

Not a fit: organizations with strict compliance rules on data location. The privacy policy states that personal data may be stored on servers outside the user's country, and the terms allow inputs and outputs to be used, after de-identification, to improve the service unless the user turns the setting off. Teams that need a private deployment but lack large GPU resources will also find it hard to run a 552B-parameter model themselves.

Not a fit: production systems that need contractually guaranteed availability. The status page shows availability between 99.47% and 99.84% across services from July to October 2026, and prices double at peak hours. Inference: systems with hard availability targets should also connect a second model provider as a backup.

How it compares

Among open-weight models, DeepSeek stands out for permissive weight licensing combined with very low prices on its own API. As of 2026-10-26, V4.1-Flash on Hugging Face has 1,640,989 downloads and 4,734 likes, and V4.1-Flash-1017, uploaded on October 17, already has 85,275 downloads and 2,920 likes. Both weights and code can be used commercially, which is more permissive than many models released for research use only.

Inference: compared with the Kimi API, both follow pay-as-you-go billing with no subscription plan. Kimi's pricing notes list a 1,048,576-token context for the K3 series, on par with DeepSeek's 1M. DeepSeek also prices peak and off-peak hours separately, with off-peak at half the peak rate. Kimi's notes add that the older K2 series will no longer be available to newly registered users from October 31, 2026. Per-token prices are best estimated from your own input/output ratio and cache-hit rate, not from list prices alone.

For self-hosted inference, DeepInfer is DeepSeek's own framework; as of 2026-10-26 it has 22,275 stars, supports only V4.1-Flash, and targets NVIDIA SM100 and Huawei Ascend 950 hardware. Its README says it does not yet include a KV cache system, prefill/decode disaggregation or load balancing. The general-purpose frameworks vLLM (95,457 stars) and SGLang (38,147 stars) are both Apache-2.0.

Inference: for teams building their own service, the general frameworks are more mature until DeepInfer adds KV caching and load balancing; DeepInfer suits teams that want to experiment with the latest official optimizations.

Getting started

Sign-up and top-up. Register on the API platform, create an API key and top up; usage is deducted from the balance. When both granted and topped-up balances exist, the granted balance is used first. The web chat and app only need an account, not a key.

First call. With the OpenAI SDK set base_url to https://api.deepseek.com, with the Anthropic SDK set it to https://api.deepseek.com/anthropic, and use the model name deepseek-flash. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the underlying models are retired; requests are served by V4.1-Flash-1017 and billed at the Flash price.

Agent tools. The docs have separate setup steps for Claude Code, Codex, OpenCode and others. DeepSeek Harness is the official desktop agent app, available for macOS (Apple silicon) and Windows (64-bit) as a preview; as of 2026-10-26 its latest GitHub release is the prerelease v0.2.1-rc.3 from October 23, 23:26 (UTC+8), with no stable release yet.

Self-hosting. Weights are downloadable from the deepseek-ai organization on Hugging Face. The DeepInfer README lists support for NVIDIA SM100 and Huawei Ascend 950. Inference: the 552B-parameter weights are very large and need multi-GPU servers, so calling the API or a third-party hosted service is more realistic for small teams.

Pricing & license

Billing. The DeepSeek API meters tokens and deducts from a prepaid balance; it has no monthly or annual plan. The web chat and app only need an account, and API fees come from the platform balance, so the two are billed separately.

deepseek-flash prices. Per million tokens: cache-hit input $0.003 off-peak and $0.006 peak; cache-miss input $0.15 off-peak and $0.3 peak; output $0.6 off-peak and $1.2 peak, per the DeepSeek pricing page (read 2026-10-26). The yuan prices are 0.02/0.04, 1/2 and 4/8 yuan, per the DeepSeek Chinese pricing page (read 2026-10-26). The concurrency limit is 2500.

deepseek-v4-pro prices. Per million tokens: cache-hit input $0.022 off-peak and $0.044 peak; cache-miss input $0.66 off-peak and $1.32 peak; output $1.98 off-peak and $3.96 peak, per the DeepSeek pricing page (read 2026-10-26). The yuan prices are 0.15/0.30, 4.5/9.0 and 13.5/27.0 yuan. The concurrency limit is 500.

Hours and search billing. Peak hours are Monday to Friday 09:00–12:00 (UTC+8) and 14:00–18:00 (UTC+8), excluding Chinese public holidays; all other hours use the off-peak price, which is half the peak price. Search and Fetch each cost $0.75 per 1,000 successful requests, or 5 yuan; searches made through server-side tool calls in DeepSeek Harness, Claude Code and Codex are free for now and billed per call from November 1, 00:00 (UTC+8).

License. Model weights are MIT, DeepInfer is Apache-2.0, DeepSeek Harness is MIT, and low-level libraries such as DeepEP, DeepGEMM and FlashMLA are MIT. The web chat, app and API service are closed source.

Risks & limits

Model names and the models behind them change. The September 10 release note says that from September 14, 12:00 (UTC+8), deepseek-v4-pro requests route to V4.1-Flash at Flash prices until V4.1-Pro launches, while the pricing page still lists deepseek-v4-pro as V4-Pro-0813 with separate prices. Apps that depend on fixed model behavior should follow the release notes and rerun their own evaluations after each upgrade.

Data use. Section 4.3 of the terms of use (updated March 27, 2026) says that inputs and outputs may be used, after de-identification, to improve the service, and that users can opt out by turning off "Improve the model for everyone"; the terms also let users train other models on outputs, including by distillation. The privacy policy (updated February 10, 2026) says data may be stored on servers outside the user's country and lists a right to refuse training use.

Security incident. The security firm Wiz published a report on January 30, 2025 04:47 (UTC+8), saying it had found a publicly accessible DeepSeek ClickHouse database holding more than a million log lines, including chat history, secret keys and backend details; the report says DeepSeek promptly secured the exposure after being notified.

Availability and peak pricing. The status page shows availability from July to October 2026 of 99.74% for the V4 Pro API, 99.47% for the V4.1 Flash API and 99.84% for the chat service. Peak prices are twice the off-peak prices, so batch jobs that cannot move to off-peak hours cost double.

The terms of service restrict use in some regions and for some purposes; see the terms for details.

How it got here

Beginnings. The deepseek-ai GitHub organization was created on October 18, 2023 17:17 (UTC+8). Two days later, on October 20, the DeepSeek-Coder repository was created; as of 2026-10-26 it has 24,311 stars and is MIT. In the years since, the company has kept a rhythm of release notes plus open weights, and the news section of the API docs has logged each model update since July 2024.

Breakout. DeepSeek-V3 was released on December 26, 2024; its repository has 104,598 stars as of 2026-10-26. The reasoning model R1 followed on January 20, 2025, with a release note stating that code and models are MIT and may be distilled and used commercially; the R1 repository has 91,957 stars as of 2026-10-26. On January 30 Wiz published the database exposure described above. V3.2 was released on December 1, 2025.

The V4 series. On April 24, 2026 the V4 preview went live and was open-sourced: V4-Pro with 1.6T total and 49B active parameters, V4-Flash with 284B total and 13B active, and context raised to 1M; the legacy names deepseek-chat and deepseek-reasoner stopped working after July 24, 23:59 (UTC+8). On August 13 V4-Pro reached general availability, adding low, high and max reasoning effort and a native Responses API, along with peak and off-peak pricing that took effect on August 17, 00:00 (UTC+8). On August 21 the experimental multimodal model V4-Flash-Vision-Exp launched with a free Files API. DeepSeek Harness 0.1.1 shipped the same day; the Harness repository was created on August 13, 19:56 (UTC+8).

V4.1. V4.1-Flash was released on September 10, retiring V4-Flash and V4-Flash-Vision-Exp, with new prices from September 10, 12:00 (UTC+8). The DeepInfer repository went public on October 16; the last commit on its main branch is from October 25, 19:59 (UTC+8), a fix to how boolean environment variables are parsed. V4.1-Flash-1017 was released on October 17 with weights uploaded to Hugging Face, and the Search and Fetch APIs opened to all users. On October 23 Harness published the prerelease v0.2.1-rc.3; the last commit on its master branch is from October 23, 19:48 (UTC+8).

Our verdict

With permissively licensed open weights and very low API prices, DeepSeek has turned high-performance large models into infrastructure ordinary developers can afford. It ships fast: the model behind a given name has changed several times within a few months, and prices have changed with it.

Recommendation: teams building agents or long-document workflows can use deepseek-flash as the main model for cost-sensitive work and schedule batch jobs off-peak. For data-sensitive cases, prefer self-hosting the open weights, or turn off the training-use setting before calling the API and review the privacy policy's data storage terms.

Inference: DeepSeek's business model uses open weights to grow its ecosystem and earns money from its own API, Harness and search endpoints. Publishing DeepInfer and Harness shows it extending from a model company into a full developer platform, but both are early, and Harness still has no stable release.

Not a fit: enterprise purchases that need contractual service guarantees, fixed model versions or a clearly stated data location. Until an enterprise agreement is offered, it is better used as one provider in a multi-vendor setup than relied on alone.

Why it matters

DeepSeek releases large models as open weights, with technical reports on how they were trained.

Sources

Official

  1. DeepSeekdeepseek.com · Accessed Oct 26, 2026
  2. Your First API Call | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  3. Models & Pricing | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  4. Models & Pricing (Chinese) | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  5. DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  6. DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  7. DeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026
  8. DeepSeek-V4-Pro GA Release | DeepSeek API Docsapi-docs.deepseek.com · Accessed Oct 26, 2026

Updates & reading

Long readEditorial

DeepSeek wants an agent you don't have to manage session by session

The DeepSeek Harness team has written a long article on X about designing an agent that stays online, keeps an eye on chats and code platforms on its own, and moves several tasks at once. It describes architectural ideas rather than shipped features, but plenty of it can be borrowed today.

Long readEditorial

A Tougher AI Leaderboard Leaves More Than Open-Weight Models Behind

Artificial Analysis says its Intelligence Index v5 widened the gap between open-weight models and the leader to 12 points from 3, and critics say a test that is 60% private can't be checked. The firm's own per-test scores show that many closed models fell alongside open-weight ones on the new tests, and that more than half of DeepSeek's deficit to the top models comes from public benchmarks anyone can rerun.

Long readEditorial

DeepSeek's Engine Was Matched in Days. The Race Moves to What It Kept.

DeepSeek released DeepInfer, the engine behind its own service, and within days vLLM and SGLang each published their own tests saying they had matched it. Engine speed is starting to look like a commons. Caching, load balancing, operations and price are where providers will now pull apart, and for most small teams the gain will arrive as more and cheaper APIs, not as something they run themselves.