How to Use DeepSeek V4

You can use DeepSeek V4 three ways: the free web chat and app, the prepaid developer Application Programming Interface (API), or self-hosted open weights under the Massachusetts Institute of Technology (MIT) license.

At Layer3Labs, we help teams integrate open-weights and commercial models into internal workflows, and we evaluate deployment paths based on latency, reliability, and token bills. The choice between chat, API, and local weights depends entirely on whether you need quick conversational answers, programmatic software integration, or complete infrastructure isolation.

Model identifiers and configurations differ across these surfaces. The default model, DeepSeek-V4.1-Flash, uses the API identifier deepseek-flash, while the larger Mixture of Experts (MoE) model DeepSeek-V4-Pro-0813 uses deepseek-v4-pro. DeepSeek reversed its plan to retire V4 Pro within four days of announcing it, so developers should verify model availability on the official documentation before locking down production pipelines.

Access Options and Model Identifiers for DeepSeek V4

Accessing DeepSeek V4 requires choosing between the official web chat interface, developer API endpoints, third-party model routers, or local deployment using open weights.

DeepSeek maintains two primary active models in the V4 family. The newer DeepSeek-V4.1-Flash model runs under the API model ID deepseek-flash. It features a 1,000,000-token (1M) context window, a maximum output limit of 384,000 (384K) tokens, native multimodal vision capabilities, and a continuous reasoning-effort setting. The larger DeepSeek-V4-Pro model runs under the API model ID deepseek-v4-pro, offering a 1M context window and native support for OpenAI Codex workflows via the Responses API, but it does not support visual inputs.

Legacy model identifiers route dynamically on the server. Requests sent to deepseek-v4-flash and deepseek-v4-flash-vision-exp automatically forward to DeepSeek-V4.1-Flash. You can review the underlying architecture and weights directly on Hugging Face or test queries in the web interface.

• deepseek-flash: points to DeepSeek-V4.1-Flash (552B total parameters, 1M context, 384K output, native vision support, reasoning-effort control).
• deepseek-v4-pro: points to DeepSeek-V4-Pro-0813 (1.6T total parameters, 49B active, 1M context, 384K output, text only, native Responses API support).
• deepseek-v4-flash: deprecated identifier that temporarily forwards traffic to deepseek-flash.
• deepseek-v4-flash-vision-exp: deprecated experimental identifier that temporarily forwards traffic to deepseek-flash.

How to Use DeepSeek V4 in Web Chat and Mobile Apps

The fastest way to use DeepSeek V4 without entering payment details or writing code is through the official web chat at DeepSeek Chat or the official DeepSeek mobile application.

The web interface offers two modes, Instant and Expert. Expert is the slower mode that reasons before answering; use Instant for quick drafting and Expert for logic, math, and code problems.

DeepSeek does not charge a consumer subscription fee or sell seat licenses for conversational access. Both the web interface and mobile applications are free to use.

• Visit chat.deepseek.com and register with an email address or mobile phone number.
• Toggle between Instant mode for rapid responses and Expert mode for deep reasoning tasks.
• Paste text or code into the prompt bar; DeepSeek does not document image upload for the chat, so test it before relying on it.
• Install the official DeepSeek app on iOS or Android for the same free access.

Setting Up the DeepSeek V4 API and Developer Endpoints

Programmatic access to DeepSeek V4 runs through a prepaid pay-as-you-go balance configured on the official developer platform at DeepSeek Platform.

DeepSeek does not offer monthly software subscriptions or fixed enterprise seat plans for API usage. Developers create an account, generate an API secret key, and top up a prepaid balance. Calls deduct credit based on exact token usage.

The endpoint supports two standard integration protocols: the OpenAI ChatCompletions format and the Anthropic API format. You can integrate DeepSeek V4 into existing software by changing the base URL to the one listed in the DeepSeek API docs, supplying your DeepSeek API key, and setting the model parameter to deepseek-flash or deepseek-v4-pro. DeepSeek-V4.1-Flash also supports an integer reasoning-effort setting from 1 to 100, which lets developers calibrate the balance between inference speed and reasoning depth.

• Authentication: Generate a private API key inside the API Keys menu at platform.deepseek.com.
• Base URL: Use the base URL listed in the DeepSeek API docs for standard ChatCompletions calls.
• OpenAI Compatibility: Use standard OpenAI software development kits by reassigning the base_url and api_key parameters.
• Anthropic Compatibility: Format headers and payloads using Anthropic specification standards documented on the DeepSeek developer portal.
• Reasoning Control: Pass an integer between 1 and 100 in the reasoning-effort setting on deepseek-flash requests; the exact parameter name is in the DeepSeek API docs.

Managing Token Costs with Cache Hits and Off-Peak Windows

DeepSeek prices API requests based on input cache status, output volume, and the specific time of day a call reaches the server.

For deepseek-flash, peak rates are $0.006 per 1M tokens for cached inputs, $0.30 per 1M tokens for uncached inputs, and $1.20 per 1M tokens for outputs. During off-peak windows, these rates drop by 50 percent to $0.003 for cached inputs, $0.15 for uncached inputs, and $0.60 for outputs. The model supports a concurrency limit of 2,500 simultaneous requests.

For deepseek-v4-pro, peak rates are $0.044 per 1M cached input tokens, $1.32 per 1M uncached input tokens, and $3.96 per 1M output tokens. Off-peak rates drop to $0.022 for cached inputs, $0.66 for uncached inputs, and $1.98 for outputs, with a concurrency ceiling of 500 simultaneous requests. The peak billing window covers 01:00 to 04:00 and 06:00 to 10:00 Coordinated Universal Time (UTC), Monday through Friday. Everything outside that window bills at off-peak rates, though developers should confirm current hours on DeepSeek API Docs as windows have changed previously.

• Context Caching: Repeated prompt prefixes automatically receive cache discounts, cutting input costs on deepseek-flash by up to 98 percent compared to uncached inputs.
• Off-Peak Scheduling: Batch jobs scheduled outside peak UTC hours save exactly 50 percent on both input and output token line items.
• Output Token Budgets: Set max_tokens limits in your request payloads to prevent open-ended reasoning chains from exhausting your prepaid balance.
• Workload Arithmetic: Processing 2M input tokens (half cached, half uncached) and 500K output tokens on deepseek-flash costs $0.453 off-peak ($0.003 for 1M hits, $0.15 for 1M misses, $0.30 for 500K outputs) versus $0.906 at peak ($0.006 + $0.30 + $0.60). This is our calculation from the published rate card.

Third-Party Hosting on Amazon Bedrock and OpenRouter

Organizations that want DeepSeek V4 inside an existing cloud account and billing relationship can access it through third-party hosts.

Both Amazon Web Services (AWS) Bedrock and OpenRouter list DeepSeek models in their catalogs. These services let engineering teams invoke DeepSeek models using existing cloud vendor billing accounts, compliance environments, and unified routing frameworks.

Pricing, throughput quotas, and regional availability on third-party hosts are governed by those platforms directly rather than DeepSeek's official rate sheet. Third-party hosts set their own prices, which are not on DeepSeek's rate sheet, so verify pricing on each provider's portal before launching large production jobs.

• AWS Bedrock: Integrates with existing AWS Virtual Private Cloud environments, Identity and Access Management policies, and enterprise billing contracts.
• OpenRouter: Lists DeepSeek models at its own per-token prices; see OpenRouter's DeepSeek page.
• Pricing Variance: Third-party hosts set their own per-token rates, which are not on DeepSeek's rate sheet; check whether caching and off-peak discounts apply on each host.

Ways to Deploy DeepSeek V4 Locally from Open Weights

DeepSeek publishes full model weights under the permissive MIT license on Hugging Face, enabling private local deployment on dedicated hardware.

High-performance serving frameworks such as vLLM and SGLang provide native inference engines for DeepSeek architectures. For desktop setups and consumer workstations, quantized model weights are distributed through community builds compatible with Ollama, llama.cpp, and LM Studio.

The architecture introduces specific memory mechanics. DeepSeek-V4.1-Flash uses a Causal Encoder-Decoder layout with 8B active input parameters and 16B active output parameters out of 552B total. It incorporates 196B sparsely accessed Engram memory parameters and 384 routed experts per MoE layer, with 6 experts activated per token alongside 1 shared expert. Its Key-Value (KV) cache requires 890 bytes per token, reducing High Bandwidth Memory (HBM) by three-quarters and Solid-State Drive (SSD) storage by seven-eighths compared to the prior generation. DeepSeek does not publish minimum hardware sizing in gigabytes, so benchmark quantized files on your specific target server before ordering hardware. Our walkthrough on how to run DeepSeek locally details the quantization and engine setup.

• Enterprise Inference: Run unquantized or FP8 weights using vLLM or SGLang on multi-accelerator servers for production API throughput.
• Local Desktop Testing: Use Ollama or LM Studio to load quantized GGUF weights on unified memory workstations.
• Containerized Stacks: Deploy Docker Model Runner images for isolated on-premise environments.
• Licensing Terms: Commercial deployment, fine-tuning, and weight modifications are permitted under the MIT license without royalty payments.

Common Setup Mistakes and Model Limitations

Several architectural constraints and naming quirks cause deployment failures for teams adopting DeepSeek V4 for the first time.

The most common mistake is passing image inputs to deepseek-v4-pro. DeepSeek-V4-Pro is strictly a text-and-code model. Computer vision, document layout extraction, and graphic comprehension require deepseek-flash.

Billing misunderstandings also disrupt integrations. Developers frequently search for monthly plan upgrades on platform.deepseek.com, only to discover that requests stop once the prepaid balance is spent. In addition, model availability remains volatile. DeepSeek announced the retirement of V4 Pro on September 10, 2026, then rescinded it within four days in response to user demand. Teams must pin specific model versions, avoid relying on temporary aliases like deepseek-v4-flash, and monitor the official change log regularly.

• Invalid Multimodal Calls: Attempting image parsing on deepseek-v4-pro fails; route vision prompts to deepseek-flash instead.
• Zero Balance Halts: Requests stop once the prepaid balance is spent, because the API has no post-paid plan.
• Hardcoded Deprecated Aliases: Reliance on legacy aliases deepseek-v4-flash or deepseek-v4-flash-vision-exp risks silent breaks when fallbacks terminate.
• Catalog Volatility: Rapid changes to model availability and off-peak windows require pin-pointed model strings in application environment variables.

Choosing the Right Deployment Path for DeepSeek V4

Selecting an access method depends on data privacy boundaries, technical capacity, and predictable request volume.

For non-technical staff and individual contributors, the free chat interface at chat.deepseek.com needs no setup. For engineering teams building scalable software, the official API with deepseek-flash offers the lowest cost per token, high concurrency limits of 2,500, and full multimodal flexibility. For organizations handling sensitive client data subject to strict confidentiality duties, hosting open weights locally or inside a private cloud removes third-party data transmission risks entirely.

DeepSeek V4 is not suitable for organizations that require strict domestic data residency without self-hosting, teams that cannot operate prepaid card billing without traditional net-30 invoicing, or teams needing turn-key enterprise agent builders. Those teams should evaluate managed commercial alternatives such as OpenAI GPT-6 Astra (see our GPT-6 Astra pricing guide) or proprietary models hosted on Amazon Web Services (AWS) Bedrock. Our recommendation would flip toward commercial managed APIs if DeepSeek eliminates off-peak pricing discounts, lowers its concurrency limits, or introduces mandatory platform seat fees. To begin evaluating how to use DeepSeek V4 for your production stack, top up a small prepaid balance at the DeepSeek platform, test your prompts against deepseek-flash using the OpenAI-compatible endpoint, and measure cache-hit ratios during off-peak hours before committing to self-hosted infrastructure.

What you need to run DeepSeek V4 yourself

DeepSeek V4 is a frontier-scale Mixture-of-Experts model, so "running it yourself" is a real infrastructure decision — not something a single laptop or gaming GPU can do. Match the path below to how seriously you need to self-host. For most teams the API or rented GPUs are the right answer; buying hardware only pays off at steady, high volume or when your data can never leave your walls.

Once DeepSeek V4 is running, the fastest way to put it to work day to day is inside Cursor — point it at the model through OpenRouter as a custom model. And if you would rather run a model on one affordable box, see Best mini PCs for local AI and Local AI hardware calculator.

NVIDIA RTX 6000 Ada (48GB)

Power users and small clusters that want cards they own

View on Amazon →