01What DeepSeek V4 Flash Is and How It Works

DeepSeek V4 Flash is a new AI model from DeepSeek, built specifically for AI agents. It went live just a few hours ago, and it's a major upgrade from the previous Flash preview version. The key change is in the training, not the model size or architecture.

The model uses the exact same architecture and size as the preview, so it runs at the same speed and cost. But it's much better at agentic tasks, like planning, coding, and using tools. It can handle up to a million tokens of context, which means it can run longer coding loops and finish more work before you need to step in.

In this episode, Julian Goldie shows how he tested it on his own benchmark, GoldieBench, and how you can use it too.

02Main Features and Benchmark Improvements

DeepSeek V4 Flash is designed for modern AI agents. It slots straight into the tools you already use, like Hermes or OpenCode. The official API is now live, and it's a big step up from the preview model.

Here are the key benchmark scores from the announcement:

  • SWE-Bench: jumped from 61.8 to 82.7
  • NL2Repo: from 39.4 to 54.2
  • Cybergym: from 38.7 to 76.7
  • DeepSWE: from 7.3 to 54.4

These scores show it's now close to models like Opus 4.8, though not at the level of frontier models like Fable 5. It also uses 12% less tokens than the previous version, so it's more efficient. The model name is DeepSeek V4 Flash 0.7.3.1, and open weights are expected soon, which would make it the second highest open weight score behind Kimi K3.

03How to Set Up DeepSeek V4 Flash on Your System

Setting up DeepSeek V4 Flash is simple. You need to have access to the official API, and then you can plug it into your existing tools. Julian shows two ways to do it: using it as a coding tool or inside the Hermes agent.

  1. Get the official API key from DeepSeek. Make sure you're using the latest version, not the preview model.
  2. If you use a tool like OpenCode, switch the model setting to DeepSeek V4 Flash.
  3. If you use Hermes, open your terminal and type hermes model to change the model.
  4. You can also use it with a custom system like AgentOS, where it's already integrated.

That's it. You can start using it for agentic tasks right away.

04How to Use DeepSeek V4 Flash with Examples

Julian tested DeepSeek V4 Flash by building over 50 things with it in a few hours. He shows games like a parachute drop game and a 3D game that works smoothly. These tests are not just for fun—they help you feel the model's logic, planning, UI design, and speed.

In the AgentOS, DeepSeek V4 Flash is added in two ways. First, there's a DeepSeek Coder tab where you can ask it to build something and see a live preview. You can then save your builds to the workspace and view them later.

Second, you can use it with Hermes. Julian created a separate profile for the model, so he can switch between models easily. This is the recommended way to use it for real tasks like automations, landing pages, websites, and agents.

05Results and Comparisons from GoldieBench

On GoldieBench, DeepSeek V4 Flash performed well. Julian notes that not every build worked—one got stuck—but overall it was fast and produced clean results. It's not at Fable 5 level, but it's close for a fraction of the cost.

Compared to other models, it's much cheaper than Fable 5, GPT 5.6, or Sonic 5. It's 60% lower price than GPT 5.6 (Luna Max) with comparable intelligence. It's within one point of GLM 5.2 and only seven points behind Kimi K3.

It's also in line with Gemini 3.6 Flash.

On the artificial intelligence analysis index, it leapfrogged many models to a mid-tier position. It's not frontier, but it's not far. The token usage is 12% lower, which means you get more for less.

06Tips, Limits, and Mistakes to Avoid

One big mistake is using the old preview model. Always make sure you're on the latest API version. The upgrade only applies to the API, not the web or app.

So if you go to DeepSeek.com, you won't see the improvements—it's still the old version there.

Also, don't expect frontier-level performance. It's great for agentic work, but not for flagship coding tasks. V4 Pro is still coming, and Julian thinks that could be at Fable 5 level.

Another tip: don't chase every new model. Have a good system where you can plug in and swap models quickly. That's why Julian has an AgentOS with separate profiles for each model.

Also, test the model yourself—benchmarks are useful, but real tests matter more.

07Who Should Use DeepSeek V4 Flash and Why

DeepSeek V4 Flash is perfect for anyone building AI agents, whether you're technical or not. Julian is non-technical and uses it easily. It's designed for automations, landing pages, websites, and coding tasks that involve agents.

It's a great choice if you want a balance of speed, cost, and performance. It's much cheaper than frontier models, and with open weights coming soon, you'll have even more flexibility. If you're already using tools like Hermes or OpenCode, you can start using it today with minimal setup.

In the AIProfitBoardroom, members have over 205 pages of wins and testimonials, showing that non-technical people can get great results with these tools. So if you're curious, give it a try.