DeepSeek V4 for web design

DeepSeek V4 arrived as a preview on April 24, 2026, in two sizes, V4-Pro and V4-Flash, and the line has moved twice since. DeepSeek V4.1 Flash replaced both on September 10, a revision called V4.1 Flash 1017 took over the API on October 17, and requests to deepseek-v4-pro have been routed to the Flash model since September 14, until a V4.1 Pro ships.

On Design Arena's Website board at the time of writing, DeepSeek V4.1 Flash scores 1305, 20 points under Claude Opus 5.5 and one point under Claude Fable 5.1. It does that with open MIT weights and an average of about two minutes per site, a combination that matters more for a production workflow than the last few Elo points.

The short call:

• The DeepSeek to use is V4.1 Flash, called as deepseek-flash. V4-Pro scores 74 points lower on the same board.
• Pages you regenerate often, such as product one-pagers, campaign heroes and variants, suit a fast model with open weights.
• 3D-heavy or agent-built apps are where the flagships still pull ahead.
• A page that stops looking generated comes from the brief. Thanor writes one per design.

What DeepSeek V4 is now#

DeepSeek's own release notes give the sequence. The search term is still "DeepSeek V4", but the model you reach today is the third release in the family.

The V4.1 Flash release describes it as the smallest model in a new architecture family, with 8B parameters active on input and 16B on output. DeepSeek says it scores ahead of V4-Pro on its benchmarks, and that tests by multiple parties put it ahead on cost, speed and runtime too.

The current weights are on Hugging Face under the MIT License, with a 1M token context window on the model card. That license is the practical difference for agencies: a team can host the model itself, on its own hardware, for client work, with no custom terms to read.

DeepSeek's pricing has peak and off-peak rates, with off-peak at half the peak rate, according to the September 10 notes. Check the current figures on DeepSeek's pricing page before you plan a large batch.

How DeepSeek scores on the design boards#

Design Arena lists several DeepSeek entries, which makes the generational change easy to see. Figures were read on November 9, 2026; Arena's WebDev board carries a November 8 date.

The September build and the October revision trade places by a few points depending on the board, which is inside the noise. The jump from V4-Pro is not: 74 points on the Website board means V4.1 Flash is expected to win about 60 of 100 votes against it, a clear verdict from voters that the newer, smaller model builds better pages.

On the Overall Agentic board the 1017 revision is fifth at 1304, but on only 249 battles, so treat that one as provisional until the count grows.

Speed is where it stands out. V4.1 Flash averages 138 seconds per site on the Website board, against 309 for Claude Opus 5.5 and 1,174 for Kimi K3.1 Pro. A fast model suits the pages you regenerate most, and shop one-pagers are the clearest case; the e-commerce landing page prompt covers what such a page has to contain. These three show what one looks like when every decision is in the brief:

What DeepSeek says it builds#

DeepSeek's October 17 notes carry the headline "Wish-Driven Programming Made Real" and a section on building complete projects from scratch. The examples are all whole builds:

• a 3D model of the Forbidden City
• a web-based racing game
• a Portal 2 recreation in the Godot engine
• a 3D scene tracing the ages from prehistory to the modern era

DeepSeek attributes the gains to reinforcement learning with self-feedback on end-to-end project development. It also says V4.1 Flash 1017 is tuned for its own DeepSeek Harness and for coding tools such as Tencent's CodeBuddy and WorkBuddy.

What the notes do not contain is any claim about visual design. There is no sentence about typography, layout or interface quality, and no design benchmark. The demonstrations are GIFs of finished builds, with no method or metric attached. That is honest, and it tells you where the design has to come from.

A model sold on building whole projects will build whatever the prompt describes, and decide whatever it does not. Ask for "a modern bakery site" and the open slots get filled from the same median every model reaches for:

• a neutral geometric sans for every heading and paragraph
• a gray sitting almost on the neutral axis
• a fade and translateY(20px) at about 300ms on every section alike

That median is why generated sites look alike, and it has nothing to do with which lab trained the model. Naming the trade more precisely does not move it either: a prompt named after an industry settles what goes on the page, not how it looks.

The pillar on model choice makes the same point across all the flagships: the six decisions that change the page are yours whichever model you pick. With an open model the point is sharper, because the cheapest way to get a better page is a better brief rather than a pricier model.

A fast model needs a complete brief#

A fast model lets you run five variants where a slow one runs one. That only helps if every variant starts from the same decided values; otherwise you get five medians.

Paste the values every time#

Each request carries the full token block, not a reference to an earlier message. A fragment written for this article:

palette   ground #F3E6D4, ink #2A1A12, accent #D9381E for the add-to-bag pill only
type      display Bricolage 120px at 0.86, body 17px at 1.55, ratio 1.3
motion    lift cubic-bezier(0.34, 1.45, 0.64, 1) over 420ms; cards stagger 70ms
product   one cut-out per card at 4:5, the hero product on its own 6s loop
order     hero, menu cards, our process, quiz, footer
refuse    no gray, no gradient, no stock photography, no carousel

Point at real media#

A cut-out bottle or a turning cup cannot be generated by a text model. Give the URLs of the films, cut-outs and stills the page uses, or the model has to improvise a placeholder. Thanor prompts carry hosted media for exactly this reason, and looping clips from the animated backgrounds page can be pasted in by URL.

Check the values survived#

Open the result and search the stylesheet for each hex and each curve in the brief. If one is missing, the model replaced it with a default, and you know which line to restate. The Gemini 3 write-up lists the drifts to expect on a long page, and they apply here unchanged.

Running DeepSeek inside the tools you already use#

DeepSeek's API is easy to slot into an existing workflow. The April notes say it accepts both the OpenAI ChatCompletions and the Anthropic API formats, so tools built for either can point at it by changing the base URL and the model name. DeepSeek names Claude Code, OpenClaw and OpenCode among the agents it works with, so the workflow in building a website with Claude Code carries over to a DeepSeek backend.

Three practical notes for a design build:

1. Use deepseek-flash. The legacy deepseek-v4-flash names still answer, but DeepSeek's docs say they are retired and redirect to the same model.
2. Give it a screenshot when you have one. V4.1 Flash has native visual understanding, and the 1017 revision is on the Image to HTML board, though behind the flagships there.
3. Self-host when the client needs it. The MIT weights make an on-premises build possible, which no closed model on these boards allows.

The kind of page a fast model is best placed for is one with a single strong idea, repeated across variants. These two each hang a whole screen on one mechanism:

If you assemble pages in Claude, ChatGPT or Cursor and hand pieces to DeepSeek, the Thanor MCP pulls a design's full art direction into those sessions without the clipboard.

Where DeepSeek trails the flagships#

The boards also show where the gap is real, and it is not on the one-page site.

3D is the widest gap. Against Claude Opus 5.5 on the 3D Design board, the 1017 revision is 113 points behind, which works out to about 66 votes in 100 for Opus. If the point of the page is a turning object or an interactive scene, that difference shows, and Kimi K3.1 Pro or one of the closed flagships is the better rental.

Two smaller caveats complete the picture. DeepSeek does not appear on Design Arena's agent-built Web Apps board, so there is no public vote yet on how it handles a multi-file React build. And DeepSeek publishes no design benchmark of its own, so every design claim about it, including this article's, rests on other people's boards.

For a page whose design starts from a picture or a film rather than from 3D code, Gemini 4 Argon is the closed flagship built around that kind of input, and the head-to-head in Claude vs Gemini covers the two biggest closed lineups.

The verdict on DeepSeek V4#

DeepSeek V4, in its current V4.1 Flash form, is a credible design model a few votes under the leaders, faster per site than the Anthropic, OpenAI and Moonshot models around it, and published under MIT weights. V4-Pro is the version to leave behind.

Ranked by what changes the page:

• The brief: values, media and refusals decide whether it looks directed.
• The media you supply: a text model cannot invent your product shot.
• V4.1 Flash over V4-Pro: 74 points on the Website board.
• DeepSeek over a closed flagship: speed, self-hosting and license, against a gap of a few votes in a hundred on a single page and a wider one on 3D.

For the other open-weight option with a stronger 3D record, read Kimi K3 and K3.1 Pro for web design. For the whole field on the same boards, read Design Arena, explained.

Run one finished brief through DeepSeek#

The test that tells you most about DeepSeek is a brief with no gaps in it, because then every difference in the result is the model's.

Open Vesperion and copy its prompt; it is free with an account. Send it to deepseek-flash through whatever tool you use, and compare the page with the moving preview, value by value. The prompt carries the palette, the type, the motion and the hosted media, so a difference you spot is one the model made.

Then pick a design from the library in the category you sell, and let the fast model do what it is good at: build the decided page quickly, as many times as the client needs, from values that stay the same between runs.