Deepseek beats Qwen in Design with 10% of the token cost
Model tested
- Models: DeepSeek V4.1 Flash and Qwen3.8-Flash, head to head (both run through omp / opencode-go)
- Use case: coding - a real feature on my hackathon project (Life OS, Next.js 16 + React 19 + Tailwind v4), not a toy prompt
Prompts I used
Instead of a synthetic benchmark I gave both models a task I actually needed done: my landing page and dashboard were full of unicode glyphs and emoji stand-ins, and I had a written plan to swap them for proper Phosphor icons plus real brand favicons. I ran both in parallel using git worktrees (separate folders, separate branches off the same base commit), identical prompt except the branch name:
Create an isolated worktree for this task: run `git worktree add ../ui-deepseek -b ui-deepseek` then `cd ../ui-deepseek`. Read the plan at plans/02-phorphor-icons.md and implement it fully - replace all generic icons on the website with proper Phosphor icons and set up the favicons, following the plan exactly with no unrelated changes. Run the project's build and lint to verify nothing broke, then commit everything on the ui-deepseek branch with a clear commit message.
(Qwen got the same text with ui-qwen; its plan path was plans/02-phorphor-icons.md. Both also got two follow-ups mid-session: "make small git commits as you go with one line msgs" and later "the jetbrains mono font looks very odd, can you swap it out for something better".)
Tasks & results
Task 1 - Phosphor icon + favicon swap (from my plan file)
- Input:
plans/02-phosphor-icons.md- replace every ASCII/unicode glyph and emoji with@phosphor-icons/reactcomponents via a centralapp/lib/icons.tsxregistry (SSR entry, no client boundary), download real brand favicons intopublic/brands/, never hand-author brand SVGs, no glyph strings left anywhere. - Output (DeepSeek): 7 commits on
ui-deepseek, 33 files touched, +360 -171. Build passes in 5.1s,tsc --noEmitclean. It ranbun run lint, got 2 errors + 1 warning, then did something I didn't ask for: it checked out the base commit in a third worktree, linted that too, and proved all 3 problems pre-existed my task. That is exactly the diligence you want from an engineer. - Output (Qwen): 7 commits on
ui-qwen, same registry architecture (the plan prescribed the skeleton), build passes in 4.2s, tsc clean at the end. But it hit a real type error mid-task (imported theIcontype from the SSR entry, which doesn't export it - TS2724) and needed a fix commit. DeepSeek read the phosphor type defs before writing the registry and never hit it. Qwen also never ranlint, even though the prompt explicitly said "build and lint". - One point both earned: my prompt had the wrong plan path (typo in the filename). Both noticed, found the real file, and carried on without asking.
Both UIs render correctly and honestly look near-identical at a glance, because the design was already fixed - the job was swapping icon machinery inside it. Hero constellation with real brand marks, trigger feed, dashboard nav: all clean on both.


The one visible quality difference I found is in the Life Graph subscriptions list. DeepSeek's GitHub mark renders white and readable on the dark UI. Qwen's GitHub favicon is the dark mark on a dark background, so it effectively disappears. Zoomed:


Small thing, but this is a UI task - pixels are the deliverable.
Task 2 - Font swap (follow-up, same session)
- Input: "the jetbrains mono font looks very odd, can you swap it out for something better"
- Output (DeepSeek): picked Geist Mono, rewired the
--font-monotoken, verified in headless Chromium withdocument.fonts.check()on both routes, committed. Done in about 90 seconds. - Output (Qwen): picked IBM Plex Mono with a design rationale that fit the site better ("slab-terminal character suits the LIFE/OS label aesthetic"). It hit a next/font type error (Plex needs explicit weights), fixed it, verified computed font in the browser. It also chased down a "Last sync" hydration mismatch I hadn't mentioned and proved it existed on the base commit before its changes. Also about 2 minutes.
Different taste calls, both defensible. I slightly prefer Qwen's reasoning here.
Cost
| Item | DeepSeek V4.1 Flash | Qwen3.8-Flash |
|---|---|---|
| Input tokens | 243,064 (+10.0M cache reads) | 882 (+8.4M cache reads, +3.4M cache writes) |
| Output tokens | 47,840 | 59,234 |
| Tool calls | 131 | 214 |
| Wall time | ~22 min | ~23 min |
| Total | $0.095 | $0.848 |
My honest verdict
- What impressed me: both models one-shot a real, multi-file refactor on an actual repo and produced mergeable branches with clean commit histories. DeepSeek's verification habits were the standout - linted the base commit to prove the failures weren't its, inspected library type defs before writing code against them. Qwen did more browser verification (37 screenshots vs 30) and its self-correction on the type error was quick and clean.
- What fell short: Qwen skipped the lint step the prompt asked for, shipped one visible UI defect (invisible GitHub mark), and needed a fix commit for a type error DeepSeek avoided entirely. DeepSeek has no real blemish on this task; if I'm nitpicking, its font pick was the safer, more boring choice.
- Worth it? DeepSeek: yes, easily - it did ~20 minutes of real feature work on my repo for under ten cents and followed the brief to the letter. Qwen: it did good work, but at 0.095.
Winner: DeepSeek V4.1 Flash. Same output quality, 9x cheaper, fewer detours, and the only place the two results visibly differ (the GitHub mark) goes in its favour. Qwen3.8-Flash is not bad - it's just beaten on exactly the axes this task measures.