DeepSeek V4: Build Apps and AI Agents for Free

There's a particular kind of YouTube video that exists at the intersection of genuine discovery and aggressive optimism. Julian Goldie's recent walkthrough of DeepSeek V4 sits squarely in that territory—part honest exploration, part hype engine, and worth parsing carefully because the underlying capability is real, even if the framing occasionally outruns it.

The premise is straightforward: DeepSeek V4, the Chinese AI lab's latest flagship model, is free to access via chat.deepseek.com, and with the right prompting technique, non-coders can build functional mini-applications in minutes. Goldie demonstrates this by asking DeepSeek to generate a Pomodoro timer and a Pong game, both specified as "single HTML" files so they can be previewed directly in the browser. The results, by his account, worked on the first try.

That single-HTML trick is actually worth noting. It's not magic—it's a constraint that forces the model to produce self-contained output you can immediately run without setting up a development environment. It's the kind of practical workaround that experienced developers know intuitively but rarely bother to explain. For someone who has never opened a terminal, it's the difference between getting something functional and staring at a wall of code wondering what to do next.

What DeepSeek V4 Can Actually Do Here

The chat interface offers a few meaningful configuration choices. There's an "instant" mode for lightweight, fast outputs and an "expert" mode that runs the full V4 model with optional "deep think" reasoning—essentially turning on chain-of-thought processing for harder problems. Web search can be toggled on or off. For simple app generation, instant mode was apparently sufficient to produce a working Pomodoro timer. For the Pong game, Goldie used the expert model with deep think enabled.

The accessibility claim holds up at this level. You don't need to know what HTML means to type "create a Pomodoro timer in single HTML" and click Run. That's genuinely new territory compared to even three years ago, and it's worth acknowledging without overselling it. A Pomodoro timer is not a SaaS business. But it is a proof that the friction between "I want this" and "this exists" has collapsed for a certain class of simple tools.

Where the video gets more interesting—and more complicated—is the second half, which involves running DeepSeek V4 Flash (a lighter, token-efficient variant) through Ollama, an open-source tool that lets you run AI models locally or interface with cloud-hosted ones via a consistent API. Goldie sets up five separate terminal windows, each running a different AI coding agent—Claude Code, ChatGPT Codex, Open Code, Open Claude, and Hermes—all pointed at DeepSeek V4 Flash as their underlying model.

The vision here is a kind of parallel AI workforce: multiple agents simultaneously building a Space Invaders game, browsing the web for AI news, spinning up landing pages, while you supervise from above. It's genuinely impressive as a demo. It's also, in Goldie's own telling, somewhat chaotic.

The Honest Breakdown

To his credit, Goldie doesn't hide the failures. Claude Code breaks and demands a login that the API setup can't provide. Open Claude crashes initially before eventually working. He acknowledges one of the five agents probably won't function. "You can see this has failed right here," he says at one point, without breaking stride. "I just want to be 100% honest with you."

That honesty matters because it surfaces something the "automate anything" framing papers over: these tools require a baseline of troubleshooting tolerance. Installing Ollama, copying terminal commands, reinstalling Open Code when it doesn't work the first time—none of this is hard in an absolute sense, but it's not one-click either. Someone who freezes when they see a terminal error message is going to have a rougher experience than the video implies.

There's also a more subtle point Goldie makes that I think deserves more attention than it gets in the breathless setup: "When I tested it out directly inside the chat, and I tried to create a website, I was like, 'This is not that interesting, right? It's not that good.' But if you go to DeepSeek and you put it inside a harness—like Open Claude—then it's super powerful."

This is a meaningful observation about how AI models actually perform in practice. DeepSeek V4 as a raw chat interface is competitive but not revelatory. DeepSeek V4 as the engine inside a purpose-built coding agent is a different product. The "harness" matters: agentic frameworks that handle task decomposition, file management, iterative feedback loops, and error recovery extract capabilities that conversational interfaces don't. This isn't a quirk of DeepSeek—it applies to most capable models—but it complicates the "anyone can do this" narrative because the harness itself has a learning curve.

The Free Tier Question

The "completely free" claim that runs through the video warrants some texture. chat.deepseek.com is indeed free with no token limits on the frontend. DeepSeek V4 Flash through Ollama is free within token limits—and Goldie is appropriately transparent that exceeding those limits either costs money or requires running a model locally, which demands hardware that most people don't own. For light personal use and experimentation, the free tier is real and useful. For running five concurrent agents building non-trivial applications, you'll eventually hit a wall.

This is worth knowing upfront, not because it invalidates the demo, but because "free" in AI contexts usually means "free until it's not." Understanding where that threshold is—and what it costs beyond it—is part of making an informed decision about whether to build workflows around these tools.

The Deeper Question the Video Doesn't Ask

What's most interesting to me about this whole category of content isn't the specific tools—those will change rapidly—but the question Goldie poses almost in passing: "I would look at where do you spend your time, and then what do you need to automate based on your time? That is by far the best way to figure out what you need to code."

That's actually a sharp piece of advice buried under a lot of "boom shakalaka." The limiting factor in personal automation has never really been the technology—it's been identifying the right targets. A tool that dramatically lowers the cost of attempting automation is only as valuable as your ability to recognize what's worth automating. A Pomodoro timer is a fine demo. Building something that saves you three hours a week requires knowing which three hours are actually automatable, which ones just feel that way, and which ones you shouldn't hand off to a machine even if you could.

DeepSeek V4 appears to be a genuinely capable model made more accessible than its Western counterparts by virtue of being free. The agent ecosystem around it—Hermes, Open Code, Codex integration—is real and functional, with the caveats that real and functional things sometimes fail and require human patience to debug.

What the "automate anything" promise can't tell you is whether you've identified the right things to automate, or whether the time you spend babysitting five terminal windows costs more than the time you save. That calculus is yours to run.

---

Marcus Chen-Ramirez is a senior technology correspondent at Buzzrag. He spent eight years writing software before deciding he'd rather write about it.