DeepSeek V4 Flash 0731 Just Dropped | Run it FREE | It’s Really Insane

August 4, 2026 4 Min Read

https://youtu.be/His7EwNQViE

On July 31, 2026, DeepSeek released V4 Flash 0731. If you skim the spec sheet, you might shrug and move on, because nothing about the model’s size changed. Same parameter count, same context window, same price. That reaction misses what actually happened here, and it’s the more interesting story: DeepSeek made the model noticeably smarter without making it bigger.

Below I’ll break down what changed, then walk you through getting it running inside VS Code, step by step. No prior terminal experience needed.

What DeepSeek actually changed

Most model upgrades follow the same recipe: more parameters, more training data, a bigger context window, heavier hardware. V4 Flash 0731 skips all of that.

The model still has 284 billion total parameters, but it only activates about 13 billion of them per token thanks to a Mixture-of-Experts design. The context window is still 1 million tokens. The architecture is untouched.

What changed is the training that happens after the base model is built. DeepSeek took the April preview version and re-trained it with a much stronger pipeline, tuned specifically for coding, AI agents, reasoning, and tool use. Same body, sharper brain. Because the architecture didn’t move, anyone already calling deepseek-v4-flash through the API gets the upgrade automatically, with zero code changes.

That’s the real takeaway: a better training pass can buy you as much real-world capability as a bigger model, without the extra compute cost.

The benchmarks (with the honest caveat)

The numbers DeepSeek published are big. On DeepSWE, a coding-agent test, the old preview scored around 7. The new version scores 54. On Terminal Bench 2.1, it went from the low 60s up to about 83. DeepSeek also reports that Flash 0731 beats its own larger V4-Pro preview on several agent benchmarks, despite activating far fewer parameters.

Here’s the part I’d want you to keep in mind: those are DeepSeek’s own reported scores, run on their own harness. Vendor benchmarks always deserve a raised eyebrow until outside tests confirm them. In this case, some did. Artificial Analysis ran independent testing and saw a real jump too, roughly 10 points on their intelligence index compared to the April release. So the improvement looks genuine, not just marketing gloss.

How to run DeepSeek V4 Flash in VS Code

We’ll connect VS Code to the model using OpenCode, a coding tool that runs right inside your editor, and OpenCode Zen, which offers a free tier that currently includes DeepSeek V4 Flash. (Free model lists do change over time, so if it’s listed when you set this up, grab it.)

Step 1: Install Node.js

Go to nodejs.org and download the Windows Installer, the .msi file. Run it, and click through the default options. If Windows asks for permission, allow it. This takes a minute, and it’s the only “setup” step that touches your system.

Step 2: Install OpenCode

Open PowerShell from the Start menu, then run this command:

npm install -g opencode-ai

That installs OpenCode globally so you can use it anywhere on your machine. Give it a few seconds to finish.

Step 3: Add the VS Code extension

Open VS Code, go to the Extensions tab, and search for “OpenCode.” Install the official one, OpenCode for VS Code by SST (it’s the top result). Restart VS Code so it loads cleanly. When it reopens, you’ll see an OpenCode icon in the top-right corner. Click it to initialize.

Step 4: Connect through OpenCode Zen

In the OpenCode panel, type /connect and choose the recommended option, OpenCode Zen. It’ll point you to opencode.ai/zen to grab an API key. Open that link, log in with Google or GitHub, and you’ll land on a dashboard listing the available models. Copy your API key.

Step 5: Pick the model

Back in VS Code, paste the key and hit Enter. You’ll see the model list, so choose DeepSeek V4 Flash 0731. It’ll ask which reasoning variant you want. Pick “High” if you want it to think a little harder before answering.

Step 6: Test it

Type hello and wait for the reply to confirm you’re connected. Then give it something real. I asked it to build an interactive HTML page with a button, opened the generated file in a browser, and it worked on the first try, in seconds.

That’s the whole setup. From here you can point it at your own projects and start building.

Should you use it?

If you’re a developer already on DeepSeek’s API, the answer is easy: you’re already getting the better model, so keep going. If you’re a beginner who wants a fast, capable coding assistant without a subscription bill, the OpenCode Zen route above gets you there in under ten minutes. Just treat the flashiest benchmark claims with healthy skepticism and run your own small tests on the kind of work you actually do.

If you build something with it, I’d love to hear what. Leave a comment on the video or drop me a note, and I’ll keep testing new models as they land.

Links from this guide: