DeepSeek V4 Preview: Powerful on Paper, Mid in Real-World Coding Tests?

DeepSeek is back with a major V4 preview release, promising a new open-source flagship that rivals top closed models while staying incredibly cheap to run. On paper, the specs and pricing are impressive. In practice, though, the current preview feels more like a benchmark-optimized experiment than a true GPT- or Claude-level alternative for real-world coding and front-end work.

What DeepSeek V4 Pro and Flash Actually Are

The V4 preview introduces two new models: DeepSeek V4 Pro and DeepSeek V4 Flash. Both are fully open source under the MIT license, which is a big win for the community and developers who want maximum flexibility.

DeepSeek V4 Pro: The Flagship

DeepSeek V4 Pro is positioned as the flagship model. It uses a mixture-of-experts style architecture with a total parameter size of around 1.6 trillion and 49 billion active parameters per token. It supports a 1 million token context window, making it suitable for long documents, large codebases, and complex multi-step tasks.

According to DeepSeek, V4 Pro aims to be the top open-source model across reasoning, STEM, coding, agentic workflows, and world knowledge. The team even claims it can rival or beat leading closed models like Claude Opus and GPT variants on certain benchmarks.

DeepSeek V4 Flash: The Cheaper, Faster Variant

DeepSeek V4 Flash is the lighter, more efficient sibling. It has a total parameter size of 284 billion with 13 billion active parameters, still with a 1 million token context. The idea is to offer near-Pro reasoning on simpler tasks while being faster and cheaper to run, especially for agent-style workflows or high-volume applications.

Pricing: Extremely Cheap Token Costs

Where DeepSeek V4 really stands out is cost. The pricing is aggressively low compared to most high-end models.

DeepSeek V4 Pro pricing:

• ~$0.14 per 1 million input tokens
• ~$0.348 per 1 million output tokens

DeepSeek V4 Flash pricing:

• ~$0.03 per 1 million input tokens
• ~$0.28 per 1 million output tokens

Combined with the 1M context window and open weights, this makes V4 Pro and Flash very attractive as infrastructure models for startups, tools, and agents—at least from a cost and architecture perspective.

Benchmarks vs Reality: Are the Claims Overhyped?

DeepSeek’s own charts suggest that V4 Pro can match or beat models like Claude Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 High on certain benchmarks, especially in coding and agentic tasks. However, hands-on testing tells a different story.

In extended reasoning and real-world coding use, other models like Qwen 3.6, Kim K 2.6, GLM 5.1, and Minimax N2.7 often perform better. On the popular Arena rankings for code, DeepSeek V4 is not at the top; it currently sits behind GLM 5.1 and Kim K 2.6, which aligns more closely with practical experience.

If you want a broader context on how this release fits into the current model landscape, it’s worth pairing this with a more general breakdown like this overview of DeepSeek V4’s architecture and positioning.

Real-World Coding & Front-End Tests

The biggest gap between promise and reality shows up when you ask DeepSeek V4 to build actual interactive projects: front-end UIs, clones of popular apps, or 3D experiences. Across multiple tests, the preview model frequently feels sloppy, incomplete, or simply uncreative compared to its competitors.

OS and App UI Clones

When tasked with building a browser-based macOS-style desktop, DeepSeek V4 Pro produced a very basic layout with minimal styling and almost no attention to real macOS structure or visual details. Despite the 1M context window, it didn’t leverage that capacity to create a richer or more accurate UI. There were no proper icons, no thoughtful layout, and no real sense of design.

A similar pattern showed up with a Slack clone request. V4 Pro roughly mimicked the column layout but missed the look and feel of Slack’s interface. In contrast, models like GLM 5.1 and Qwen 3.6 Plus generated UIs that much more closely matched Slack’s visual style and structure.

SVG and 3D Generation

On SVG tasks, DeepSeek V4 Pro can produce usable output—like a butterfly illustration—but when compared side-by-side with Qwen 3.6 Plus, the difference in quality is obvious. Qwen’s SVGs tend to be more polished and better structured.

For 3D work with Three.js, the gap widens. When asked to create a 3D PS5 controller, DeepSeek V4 Pro generated a shape that barely resembled a controller. The geometry and proportions were off, and the result felt more like a random object than a recognizable device. Competing models produced something that clearly looked like a controller, with more accurate attributes and structure.

Front-End Landing Pages

On SaaS landing pages and marketing sites, DeepSeek V4 Pro often feels like an upgraded GPT-3.5 rather than a top-tier 2026 model. It can follow typography and layout instructions reasonably well and sometimes maintains consistent structure across a long prompt, but the dynamic behavior and polish are lacking.

Animations, interactions, and fine-grained styling are usually basic or missing. The code tends to work at a minimal level but rarely feels production-ready or creatively designed.

Interactive Experiences and Games

When asked to build more complex interactive experiences, the limitations become even clearer:

• Minecraft-style clone: DeepSeek V4 Pro produced a very barebones sandbox. You could move and place a few blocks, but there was no lighting, no textures, no inventory system, and little sense of polish. Other models—even the cheaper Flash variant in some tests—managed more features like infinite terrain and basic inventory.
• Off-road EV durability test: V4 Pro failed to complete the generation, while Minimax N2.7 produced a working (if imperfect) interactive demo where you could move the car and test the environment.
• Instagram feed clone: Minimax N2.7 delivered a UI that closely resembled Instagram’s feed. DeepSeek V4 Pro, on the other hand, struggled with bugs, messy structure, and incomplete output.

In short, DeepSeek V4 Pro can generate code that runs, but it often stops short of the completeness and refinement you’d expect from a model that claims to rival top closed systems.

Where It Does Better

It’s not all negative. There are cases where DeepSeek V4 Flash actually feels more responsive and practical than Pro for simpler prompts. On some landing pages and smaller UI tasks, it sticks to the brief reasonably well and "gets the job done" even if the result is still mid-tier.

For a 360-degree product viewer, V4 Pro produced a decent front-end with a 3D product and rotation controls. It wasn’t exceptional—more like a 6/10—but it was functional and visually acceptable. Minimax, by comparison, had a nicer front-end but didn’t manage the full 3D rotational behavior.

How It Stacks Up Against Other Models

When put side-by-side with the latest closed models, the preview version of DeepSeek V4 clearly lags. Claude Opus 4.7, for example, can generate far more polished, creative, and structurally sound front-end experiences with the same prompts. Even earlier versions like Opus 4.6 would likely outperform V4 Pro on many of these tasks.

Among open and regional models, DeepSeek V4 also struggles to stand out. Chinese models like Minimax N2.7 and Kim K 2.6 often deliver more complete and visually accurate clones of popular apps. GLM 5.1 and Qwen 3.6 Plus tend to produce better code structure, more realistic UIs, and richer interactive behavior.

If you’re exploring how these models compare more broadly across the ecosystem, you may also find this broader first-look analysis useful: a detailed first impressions review of DeepSeek V4.

Access, Open Weights, and Use Cases Today

Despite the mixed performance, DeepSeek V4 has real strengths that make it worth watching:

• Open weights on Hugging Face: You can download and run the models yourself, fine-tune them, or integrate them into your own stack.
• Cloud access: The models are available through providers like OlaML and via DeepSeek’s own chatbot interface, where you can switch between Pro and Flash modes for free testing.
• Massive context + low cost: For long-context tasks where perfection isn’t critical—like internal tools, log analysis, or rough prototyping—the combination of 1M tokens and ultra-low pricing is compelling.

Right now, DeepSeek V4 Preview feels best suited as a cheap, high-context workhorse for experimentation, research, or internal agent workflows, rather than as a primary model for polished, production-grade front-end or complex interactive experiences.

Verdict: Impressive Foundations, Mid Real-World Output (For Now)

On paper, DeepSeek V4 Pro and Flash are extremely impressive: huge context, clever architecture, MIT-licensed open weights, and some of the lowest token prices on the market. In practice, the current preview falls short of the hype when it comes to real-world coding, UI cloning, and 3D or game-like experiences.

The generations are often:

• Sloppy or incomplete
• Visually unpolished
• Less creative than competing models
• Prone to bugs or missing features in complex projects

That doesn’t mean DeepSeek V4 is a failure. It’s a strong foundation, especially for open-source and cost-sensitive use cases, and this is explicitly a preview. A future official release could significantly improve generation quality and bring it closer to the bold benchmark claims.

For now, though, "cheaper" doesn’t automatically mean "better." DeepSeek V4 Preview is a mid-tier real-world performer wrapped in a very attractive pricing and licensing package. If you’re building serious front-end or interactive products today, you’ll likely still lean on more mature models—but DeepSeek V4 is absolutely a project to keep an eye on as it evolves.