01What Is DeepSeek V4 and How Does It Work?

DeepSeek V4 is a new open-source AI model released by DeepSeek, available for free on their chat platform and API. It uses a mixture-of-experts architecture, meaning only a small part of the model activates for each task, saving compute.

The model comes in two versions: V4 Pro with 1.6 trillion total parameters (but activates only 49 billion at a time) and V4 Flash with 284 billion parameters (activates 13 billion at a time). Both have a 1 million token context window, allowing you to process very long documents.

DeepSeek V4 is designed to rival top closed-source models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, according to benchmarks shown in the episode.

02Main Features and What's New in DeepSeek V4

  • Open source: Downloadable and runnable locally, with quantized versions for smaller setups.
  • Million token context: Standard across all DeepSeek options, handling huge inputs efficiently.
  • Enhanced agentic capabilities: State-of-the-art in agentic coding benchmarks.
  • World knowledge: Trained on over 32 trillion tokens including web, code, and science.
  • Efficient memory: V4 Pro uses only 27% of compute and 10% of KV cache compared to V3.2; Flash uses 10% compute and 7% memory.
  • New attention methods: Compressed sparse attention and heavily compressed attention (HCA) help handle long context faster.

Also new are the thinking modes: non-think (instant), think high (step-by-step), and think max (deep reasoning, using up to 384k tokens).

03How to Get Started with DeepSeek V4

To start using DeepSeek V4, follow these steps from the episode:

  1. Go to chap.deepseek.com (the chat interface).
  2. Choose between Expert mode and Instant mode.
  3. Optionally select a thinking mode: non-think, think high, or think max (deep think is slower but better).
  4. Start typing your prompt in the chat box.
  5. For API access, go to platform.deepseek.com and grab an API key.
  6. Plug the API into AI agents or tools like Hermes or OpenClaw.

You can also download the model from Hugging Face; it may appear on LM Studio later.

04How to Use DeepSeek V4: Real Examples

In the episode, Julian tested DeepSeek V4 by creating a landing page and building a game. He used expert mode without deep think first, then tried deep think.

For the landing page, he copied his website's content, asked for a new fun alternative, and embedded a video. Instant mode responded quickly but produced a basic design. Expert mode without reasoning gave a slightly better but still dated output.

Then he tried building a computer-versus-human game with deep think. This took longer but produced a better result, though still buggy and laggy compared to GPT-5.5.

He compared outputs side-by-side with Claude Opus 4.6, which he felt was much better for coding. Julian noted that for serious coding tasks, he prefers Claude or GPT-5.5.

05Benchmarks and Real-World Test Results

On benchmarks, DeepSeek V4 Pro Max scored 57.9% on SimpleQA verified, beating Claude Opus 4.6 Max at 46.2% and GPT-5.4 at 45.3%. On coding completions, it scored 3206, the highest of any model tested. On the APEX Shortlist, it scored 90.2%, beating every other model including Gemini 3.1 Pro.

However, in real-world testing in the episode, the outputs felt dated—like version 3. Julian said, "I'm not going to tell you it's amazing if it's not amazing." He suggested that maybe not using reasoning mode hurt, but even with deep think, it was slower and not as impressive as GPT-5.5.

Flash performed on par with Pro on simple agent tasks but was faster and cheaper. The episode noted that models are releasing daily, making comparisons hard.

06Cost, Efficiency, and API Details

DeepSeek V4 is significantly cheaper to run than closed models. The episode showed a cost table with huge differences—input costs for DeepSeek were much lower than Claude Opus or GPT-5.4.

V4 Pro uses only 27% of compute and 10% of KV cache compared to V3.2. Flash uses 10% compute and 7% memory, making it very efficient.

The API is available now at platform.deepseek.com. Old versions (deepseek-chat and deepseek-reasoner) will be fully retired after July 24th, so you must move to V4.

You can also run it locally via Hugging Face or LM Studio (once available).

07Tips, Limits, and Mistakes to Avoid

  • Use deep think (think max) for complex tasks—non-think mode gives poor quality, as seen in the episode.
  • Be patient: think max takes minutes to respond, but produces better outputs.
  • Don't expect GPT-5.5-level coding; Julian found outputs dated and laggy.
  • Be aware of usage limits if you use codex with GPT-5.5; DeepSeek is more affordable.
  • The UI feels dated compared to ChatGPT; don't judge the model by the interface.
  • Watch out for bugs in generated code (like the game lagging).

Remember: benchmarks are impressive, but real-world results may vary. Test thoroughly before relying on it.

08Who Is DeepSeek V4 For and Why It Matters

DeepSeek V4 is for developers, AI enthusiasts, and businesses that want open-source access to a powerful model without high costs. It's especially useful for AI agent workflows, content generation, coding assistance, and research.

Because it's open source, you can run it locally, and quantized versions make it possible on modest setups. The API pricing is a game-changer for developers who need scale.

Julian noted that if you have the budget, Claude is still the best for coding, but DeepSeek V4 offers a solid free and cheap alternative. It matters because it brings competition and choice to the AI market.