DeepSeek V4 (Pro and Flash): Specs, Pricing, Context Window and Availability Facts

DeepSeek V4 is a family of open-weight Mixture-of-Experts large language models from DeepSeek, released as a preview on 24 April 2026 in two sizes, V4-Pro and V4-Flash. This profile collects the published facts about both: API identifiers, parameter counts, context window, output limits, peak and off-peak prices, thinking modes, availability, licence, provider-reported benchmarks and the lifecycle of each version. V4-Pro is still served on the DeepSeek API as DeepSeek-V4-Pro-0813; V4-Flash was retired on 10 September 2026 and its API name now routes to V4.1-Flash, which has its own profile. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Where DeepSeek publishes nothing on a point, the page says so.

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 and model cards as of November 2026. Model details change, so confirm against DeepSeek's current documentation.

Key takeaways

• Preview on 24 April 2026, Pro generally available on 13 August 2026. The DeepSeek API changelog records V4-Pro and V4-Flash on 24 April, a V4-Flash update on 31 July and V4-Pro general availability on the app, web and API on 13 August.
• 1.6T total and 49B active parameters for V4-Pro. The preview model card lists V4-Pro at 1.6T total and 49B activated parameters and V4-Flash at 284B and 13B, both with a 1M token context; the 0813 release card shows 1.7T in its size box.
• $1.32 input and $3.96 output per million tokens at peak. deepseek-v4-pro is billed at half those rates off-peak ($0.66 and $1.98), with cache-hit input at $0.044 peak and $0.022 off-peak.
• MIT licence for the weights. Every V4 model card states that the repository and the model weights are licensed under the MIT License.
• V4-Flash is retired; V4-Pro continues. On 10 September 2026 DeepSeek retired V4-Flash and routed deepseek-v4-flash to V4.1-Flash, and stated it would continue API service for V4-Pro after 14 September 2026.

What is DeepSeek V4?

DeepSeek V4 is the fourth generation of DeepSeek's general-purpose language models, following V3.2 (the preview card compares base models against V3.2-Base). The technical report is titled "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence" (arXiv 2606.19348, as cited on the model cards). The family has two members:

• V4-Pro: the larger model, aimed at knowledge-heavy and complex agentic work. The preview card says V4-Flash in its maximum reasoning mode trails Pro on pure knowledge tasks and the most complex agentic workflows.
• V4-Flash: the smaller, cheaper model with a far smaller activated parameter count. It was replaced on the API by V4.1-Flash in September 2026.

Both are Mixture-of-Experts models with three operating modes: non-thinking, thinking at high effort and thinking at maximum effort. The preview card describes the architecture as a hybrid of Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), with Manifold-Constrained Hyper-Connections (mHC) for residual connections and the Muon optimiser. Pre-training used more than 32T tokens; post-training trained domain-specific experts with supervised fine-tuning and reinforcement learning (GRPO), then consolidated them through on-policy distillation. For the full DeepSeek line-up, see DeepSeek Models.

Key facts at a glance

As of 2026-11-15. Source: DeepSeek API pricing page, API changelog and Hugging Face model cards.

The parameter figures differ between cards. The 0813 and 0731 cards attach a DSpark speculative decoding module to the checkpoint, and their size boxes show 1.7T and 304B; DeepSeek does not explain the difference from the preview figures on the pages checked. Use the preview card's total and active figures for architecture comparisons and the size box for storage planning.

Release and version history

As of 2026-11-15. Dates from the DeepSeek API changelog and Hugging Face model cards.

The 0813 and 0731 cards each describe themselves as the official release "superseding the preview version".

How much does DeepSeek V4 cost?

As of 2026-11-15. Per million tokens. The pricing page uses the "$" sign without naming the currency. Source: DeepSeek API pricing page.

• Thinking tokens: thinking is the default mode, so reasoning output counts towards output tokens. Budget for it at max effort, where the cards recommend up to 384K output.

Worked example at peak rates on a cache miss: 100,000 input tokens and 10,000 output tokens on deepseek-v4-pro cost $0.132 for input plus $0.0396 for output, about $0.17 in total. The same request off-peak costs about half that. For how caching changes the maths, see LLM Pricing Explained: Tokens and Caching.

Context window and output limits

Both V4 sizes have a 1M token context window. The pricing page lists a maximum output of up to 384K tokens. The model cards recommend allowing 384K output tokens for the high and max reasoning levels, and the preview card says the Think Max mode needs a context window of at least 384K tokens.

For what these numbers mean in practice, see LLM Context Windows Explained.

Modalities and tools

V4-Pro accepts text and returns text; the pricing page marks vision as not supported for deepseek-v4-pro. The retired V4-Flash was text only, and DeepSeek's experimental vision variant (V4-Flash-Vision-Exp) was retired on the same date.

• API features (deepseek-v4-pro): JSON output, tool calls, Responses API, Anthropic API compatibility and chat prefix completion (beta); FIM completion (beta) in non-thinking mode only.
• Thinking control: non-thinking and thinking modes, with thinking as the default; the reasoning_effort parameter takes low, high or max.
• Sampling: the cards recommend temperature 1.0, with top_p 0.95 for agentic scenarios and 1.0 otherwise.
• Application APIs: the 13 November 2026 changelog entry adds web search, file input, URL reading, model safety, ASR and TTS for the deepseek-flash model; it does not name deepseek-v4-pro.
• Self-hosting: the 0813 card gives serving examples for vLLM and SGLang with DSpark speculative decoding, points to an inference folder for local demos, and lists Docker Model Runner and quantised builds for llama.cpp, Ollama and LM Studio. The release has no Jinja chat template; an encoding folder provides Python scripts to encode OpenAI-format messages.

Tool calls and Anthropic-compatible endpoints let existing agent harnesses point at DeepSeek by changing a base URL and key. Treat that key as a credential with spend authority. See the LLM API Key Security and LLMjacking Guide.

Where is DeepSeek V4 available?

As of 2026-11-15. Source: DeepSeek API changelog, pricing page and Hugging Face model cards.

Access restrictions

DeepSeek V4 is not a restricted-access model. V4-Pro is listed on the public pricing page with a published concurrency limit, and the weights are downloadable under the MIT License. DeepSeek publishes no trusted-access or gated tier for V4 on the pages checked. Whether your own organisation permits DeepSeek services is a buyer-side policy question. See the DeepSeek and High-Risk AI Vendor Assessment Guide and the DeepSeek Federal Exclusion Orders and FASCSA Guide.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported from DeepSeek's Hugging Face model cards: the V4-Pro preview card (technical report published 26 April 2026) and the V4-Pro-0813 card (general availability 13 August 2026). Independent leaderboard results are not included because none were checked against their own published pages for this profile.

The 0813 card notes that its code-agent tasks ran in DeepSeek Harness (minimal mode) at max reasoning effort, temperature 1.0 and top_p 0.95, and that its DSBench rows use internal DeepSeek test sets. The cards also report results for models from other providers; those comparisons are DeepSeek's own runs and are not reproduced here. Scores in non-thinking mode are much lower on reasoning tasks (the preview card reports 7.7 on HLE). For how to read these figures, see How to Read LLM Benchmarks.

Model card and system card

DeepSeek publishes model cards on Hugging Face for each V4 checkpoint and a technical report on arXiv. It publishes no separate system card for V4 on the pages checked. What the cards state:

• Architecture and training: MoE with CSA and HCA hybrid attention, mHC residual connections and the Muon optimiser; pre-training on more than 32T tokens; expert training with SFT and GRPO, then on-policy distillation.
• Modes: Non-think for fast responses, Think High for deliberate reasoning and Think Max, which requires a special system prompt.
• Release notes in the card: the 0813 and 0731 cards attach a DSpark speculative decoding module whose draft weights come from the same checkpoint.
• Safety: none of the V4 cards has a dedicated safety or limitations section. The only limitation stated is that V4-Flash-Max trails V4-Pro on pure knowledge tasks and the most complex agentic workflows. The contact listed is [email protected].

Data handling

DeepSeek has no V4-specific data policy. Two documents apply to the hosted service:

• Open Platform terms of service: the platform is owned and operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. DeepSeek assigns to the customer any rights it has in outputs. The terms are governed by the laws of the People's Republic of China (mainland). The terms do not state whether API inputs or outputs are used to train DeepSeek models, and they do not state retention periods or storage location.
• Privacy policy (last updated 10 February 2026): covers DeepSeek apps, websites, software and related services. It states that DeepSeek collects, processes and stores personal data in the People's Republic of China, lists training and improving its models as a purpose for user input, and lists a right to opt out of training use by contacting [email protected]. It says data from applications developers build on the open platform is not covered by the policy and is the developer's responsibility. No fixed retention periods are given.

Third-party hosts of the weights set their own data terms. Self-hosting the MIT-licensed weights keeps prompts inside your own environment, which changes the risk to one of model supply chain and serving security. See the Open-Weight Model Security Guide and the Self-Hosted LLM Security Guide.

Deprecation status

As of 15 November 2026, deepseek-v4-pro is current. On 10 September 2026 DeepSeek wrote that it had "decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026", with billing unchanged and further notice promised before any change. V4-Flash was retired on 10 September 2026.

"Temporarily routed" means code still calling deepseek-v4-flash runs on a different model today and may fail when the route is withdrawn. Track dates across providers in the LLM Deprecation and Retirement Tracker.

How DeepSeek V4 differs from other models in its family

As of 2026-11-15. Source: DeepSeek pricing page, API changelog and model cards. Models listed by release date.

The documented differences are size, price (V4-Pro peak output costs 3.3 times deepseek-flash peak output), vision support, concurrency (500 against 2,500) and the 13 November application APIs, which DeepSeek announced for deepseek-flash. All share the 1M context and 384K output listing. See the DeepSeek V4.1 Flash profile, and DeepSeek vs Qwen vs Kimi vs GLM for cross-provider fields.

Practitioner checklist

• Search code and configuration for deepseek-v4-flash, deepseek-chat and deepseek-reasoner, and move each call to an explicit current model name.
• Set reasoning_effort deliberately (low, high or max) and size max output for reasoning tokens, not only the visible answer.
• Model cost with peak and off-peak windows in UTC; schedule batch-style jobs off-peak where latency allows.
• Design for the 500 concurrency limit with queuing and retry back-off before scaling agent workloads.
• Decide between the hosted API and self-hosted weights based on data residency: the hosted service is governed by PRC law and the privacy policy states processing in the PRC.
• If self-hosting, verify checkpoint hashes, pin the repository revision and isolate the inference server.
• Store DeepSeek API keys in a secrets manager, separate them by environment and set spend alerts on the account balance.
• Run your own evaluation and red-team tests, since the V4 cards publish no safety section.

For spend controls, see How to Set Spend Limits on LLM API Keys.

Standards and references

• DeepSeek: Models and pricing (accessed 2026-11-15)
• DeepSeek: API change log (accessed 2026-11-15)
• DeepSeek on Hugging Face: DeepSeek-V4-Pro model card (preview) (accessed 2026-11-15)
• DeepSeek on Hugging Face: DeepSeek-V4-Pro-0813 model card (accessed 2026-11-15)
• DeepSeek on Hugging Face: DeepSeek-V4-Flash-0731 model card (accessed 2026-11-15)
• DeepSeek: Open Platform Terms of Service (accessed 2026-11-15)
• DeepSeek: Privacy Policy (accessed 2026-11-15)

Related NHI Mgmt Group resources: DeepSeek Models, DeepSeek V4.1 Flash, DeepSeek vs Qwen vs Kimi vs GLM, Frontier AI Models Compared, DeepSeek and High-Risk AI Vendor Assessment Guide.