How to Use DeepSeek V4 Pro & Flash for Free with NVIDIA NIM

DeepSeek V4 has arrived, and you can already use its two main models—V4 Pro and V4 Flash—through NVIDIA’s NIM APIs with free developer access. That means you can test powerful long‑context, reasoning‑focused models in your own apps and coding tools without running GPUs yourself or paying per token right away.

DeepSeek V4 Pro vs V4 Flash: Which One Should You Use?

DeepSeek V4 comes in two main flavors for developers: Pro and Flash. Both are built for serious coding and long‑context work, but they target slightly different use cases.

DeepSeek V4 Pro is the flagship model. It’s a mixture‑of‑experts system with about 1.6 trillion total parameters and roughly 49 billion active parameters per request. It supports up to a 1 million token context window, making it ideal for:

• Hard reasoning and complex problem‑solving
• Deep code understanding across large repositories
• Long‑context agents and tool‑using workflows
• Document analysis and multi‑file planning

DeepSeek V4 Flash is the smaller, faster sibling. It has around 284 billion total parameters and about 13 billion active parameters, but still supports the same 1 million token context window. It’s designed for:

• Fast responses and lower‑cost inference
• Summarization and quick chat interactions
• Lightweight coding tasks and utility scripts
• Routing requests before sending hard problems to Pro

In short: use V4 Pro when you want maximum quality and reasoning, and V4 Flash when you care more about speed and efficiency. For a deeper look at the architecture and benchmarks behind these models, you can check out this breakdown of DeepSeek V4 Pro and Flash.

What Is NVIDIA NIM and Why It Matters

NIM stands for NVIDIA Inference Microservice. In practice, it means NVIDIA hosts the model for you on its GPU infrastructure and exposes a simple API endpoint you can call from your apps and tools.

The best part: the NIM API is OpenAI‑compatible. If your tool already supports OpenAI‑style providers, you can usually plug in NVIDIA with just a base URL, an API key, and a model name.

Key details:

• Base URL: https://integrate.api.nvidia.com/v1
• Chat completions endpoint: https://integrate.api.nvidia.com/v1/chat/completions
• Model names on NVIDIA NIM:
– deepseek-ai/deepseek-v4-pro
– deepseek-ai/deepseek-v4-flash

Note the difference: on DeepSeek’s own API, the models are named deepseek-v4-pro and deepseek-v4-flash. On NVIDIA NIM, you must include the deepseek-ai/ prefix. If you get the name wrong, your tool may fail to load the model.

NVIDIA describes this as free access for prototyping and development through the NVIDIA Developer Program. It’s great for testing, building demos, and experimenting—but you shouldn’t treat it as an unlimited production backend. Rate limits and terms can change, so always check NVIDIA’s current developer terms if you’re planning a serious deployment.

Step‑by‑Step: Getting Your NVIDIA API Key

You can start using DeepSeek V4 Pro and Flash via NIM in a few minutes. Here’s the basic flow:

1. Go to build.nvidia.com.
2. Search for DeepSeek V4 in the model catalog.
3. Open the model page for either DeepSeek V4 Pro or DeepSeek V4 Flash (from DeepSeek AI).
4. If a third‑party model warning appears, read it and continue.
5. On the model page, try the in‑browser playground first—send a prompt to confirm the model works for your account.
6. Click “Get API key” on the model page.
7. Sign in or create an NVIDIA account if needed. This also joins you to the NVIDIA Developer Program.
8. Copy your API key and store it securely.

Once you have the key, you’re ready to call the model from your own code or any OpenAI‑compatible tool.

Using DeepSeek V4 via the OpenAI SDK

You don’t need a special NVIDIA SDK to get started. The standard OpenAI SDK works as long as you point it at NVIDIA’s base URL and use your NVIDIA key.

High‑level steps:

1. Install the OpenAI client in your project (for example, via npm or pip).
2. Configure the client with:
– Base URL: https://integrate.api.nvidia.com/v1
– API key: your NVIDIA key
3. Call the /chat/completions endpoint with:
– model: deepseek-ai/deepseek-v4-pro or deepseek-ai/deepseek-v4-flash
– messages: the usual OpenAI‑style array of system/user/assistant messages.

The request format is the same as you’d use for OpenAI’s chat models, which makes migration or experimentation straightforward.

Reasoning Effort: Tuning Speed vs Depth

DeepSeek V4 on NVIDIA NIM supports a special parameter called reasoning_effort. This controls how much “thinking” the model does before answering.

Supported values:

• none – Disables the extra thinking process. Fastest, but shallower reasoning.
• high – Normal reasoning mode (default). Good balance of quality and speed.
• max – Strongest reasoning mode. Best for very hard tasks, but slower and more token‑intensive.

Practical suggestions:

• For V4 Flash, try none or high for fast coding help, summaries, and routing.
• For V4 Pro, use high for everyday coding and max when debugging tricky bugs, planning large refactors, or handling complex multi‑file reasoning.

The advantage is that you don’t need to switch models to change behavior—you can keep the same model and just adjust reasoning_effort per task.

Context Window and Token Limits

DeepSeek V4 Pro and Flash both support a 1 million token context window at the model level, which is huge for long‑context agents, large codebases, and big documents.

However, NVIDIA’s NIM documentation currently lists max_tokens up to 16,384 for these endpoints. That means:

• The underlying model can handle very long contexts.
• The specific host (NVIDIA NIM) may expose smaller practical limits for now.
• Your tools might also apply their own chunking, summarization, or prompt limits.

Think of 1M context as a capability of the model family, not a guarantee that every endpoint or coding tool will automatically use all of it. For real projects, always check the limits of the specific API and client you’re using.

Connecting DeepSeek V4 to Coding Tools

Because the NIM API is OpenAI‑compatible, many popular coding assistants can use DeepSeek V4 simply by adding NVIDIA as a provider.

Using Codium CLI

If you use Codium CLI, setup is straightforward:

1. Open Codium CLI.
2. Run /connect and choose NVIDIA as the provider.
3. Paste your NVIDIA API key.
4. Run /models and select DeepSeek V4 Pro or DeepSeek V4 Flash if they appear in the list.

Once connected, you can switch between NVIDIA‑hosted models exposed in the catalog without changing your workflow.

Using Other OpenAI‑Compatible Tools

The same pattern works in many tools that support custom OpenAI endpoints, such as OpenCode, Cursor, Aider, RueCode, Kite, and others.

General configuration:

• Base URL: https://integrate.api.nvidia.com/v1
• API key: your NVIDIA key
• Model: deepseek-ai/deepseek-v4-pro or deepseek-ai/deepseek-v4-flash

If the tool has a built‑in NVIDIA provider but doesn’t yet show the DeepSeek V4 models, you can usually add them manually using the OpenAI‑compatible settings.

When to Use Flash vs Pro in Real Workflows

To get the most out of these models, it helps to split tasks by speed vs depth.

Good uses for DeepSeek V4 Flash:

• Quick repository overviews or file explanations
• Small code edits and refactors
• Writing tests for simple functions
• Summarizing documentation or long text
• Generating commit messages and changelog entries
• Extracting structured data from text
• Acting as a router model before forwarding only hard tasks to Pro

Good uses for DeepSeek V4 Pro:

• Full agentic coding workflows: analyze project, plan, implement, test, and explain changes
• Debugging non‑obvious bugs that span multiple files or layers
• Implementing complex features across a large codebase
• Working with long design docs, API docs, and multiple source files together
• Any scenario where you care more about correctness and deep reasoning than raw speed

To really understand the difference, don’t just ask each model a single random question. Instead, run the same realistic workflow through both: have them implement the same feature, fix the same bug, or summarize the same document, then compare speed, quality, and how much cleanup you need. For more benchmark‑style comparisons, you may find this coding and UI showdown between DeepSeek V4 Pro and other top models helpful.

Final Thoughts and Caveats

DeepSeek has quickly become one of the most important players for people who want powerful, cost‑effective alternatives to closed models. V4 continues that trend, pushing hard into long‑context agents, coding, tool use, and serious reasoning.

NVIDIA hosting DeepSeek V4 Pro and Flash through NIM makes it much easier to try these models today. If your tool can talk to an OpenAI‑style API, you can likely plug in NVIDIA and start experimenting in minutes.

To recap the basics:

• Go to build.nvidia.com and find DeepSeek V4 Pro or DeepSeek V4 Flash.
• Get your NVIDIA API key from the model page.
• Use https://integrate.api.nvidia.com/v1 as the base URL.
• Use deepseek-ai/deepseek-v4-pro for maximum quality and deepseek-ai/deepseek-v4-flash for speed.
• Adjust reasoning_effort (none, high, max) to balance speed vs depth.

Free developer access is an excellent way to explore these models, prototype ideas, and build coding workflows. Just remember that availability, limits, and terms can change, so always review NVIDIA’s current policies before relying on this setup for production systems.