# I compared Qwen 3.8 27B vs 3.8 FN vs DS V4.1 Flash on a real frontend task (one shot). Results inside — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/3364e999d796d8db

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-17 19:59 ingested 2026-11-18 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xyujv0/i_compared_qwen_38_27b_vs_38_fn_vs_ds_v41_flash/)

## retained text



```
So last time I ran [3.8 FN against DS V4.1 Flash](https://www.reddit.com/r/LocalLLaMA/comments/1xs7ffm/same_prompt_deepseek_v41_flash_vs_qwen_38_fn_94/) on a synthetic test, which was mostly for show, no real usage beyond testing their 3D capabilities. I wanted to do something similar, but with purpose, a task that anyone might actually want to run, build a website with a refined spec and a completion checklist, but not strict enough that it removes all of the models "freedom". As last time it was a single prompt: build a single file landing page for a fake famous NYC photographer (real photos, accessible, no console errors, yadda yadda). Each model started on an empty folder w/ the same toolkit (node, headless chrome, image viewer tool). No changes were made to the files after the model was done.

**Setup:** RTX 5090 + RTX 5060 Ti, Ryzen 9 9950X, 128 GB DDR5. Qwen 3.8 27B and Qwen 3.8 FN were NVFP4. DeepSeek V4.1 Flash (1017) was an IQ1\_M quant.

**Qwen 3.8 27B:** \~12 min wall time, 81 tool calls (3 failed), 182 tok/s, 3.8 MB HTML

[Qwen 3.8 27B Website](https://preview.redd.it/f9fgj6pa782i1.png?width=3024&format=png&auto=webp&s=f7b0ddcfaf324157bb1a3239fbd5251f48aae4e0)

The simplest of all three, not very daring. A dark gallery with lime accents and 12 frames. All 3 tool failures were in its own test script (wrong file path, missing elements). It has empty gaps in the photo grid, and the photos in the bio are stretched. Decent, I'd refine it a bit but definitely usable.

**Qwen 3.8 FN:** \~12 min wall time, 124 tool calls (4 failed), 303 tok/s, 2.1 MB HTML

[Qwen 3.8 Flash Next Website](https://preview.redd.it/thqtru1k782i1.png?width=3024&format=png&auto=webp&s=5242f84803ea85a50c82a1852310667d7303d080)

My personal favorite by far, a hero where your pointer reveals a mosaic of the portfolio, 19 captioned frames with a nice blur/fade in animation, a live clock and form validation. The 4 failures were 3 rejected edits and a search sent with an empty query (intersting!). It also found and fixed bugs using its own tests.

**DeepSeek V4.1 Flash:** \~36 min wall time, 155 tool calls (1 failed), 94 tok/s, 1.4 MB HTML

[DS V4.1 Website](https://preview.redd.it/yyij8xiv782i1.png?width=3024&format=png&auto=webp&s=57a5c934480b106d500058848aa0e42e8a1cc73e)

The boldest one, the model committed to it (and it shows in wall time), cool accents all around, terrible typography. It has only 10 frames, and the hero photo covers a bit of the last "T" of the headline + toolbar logo is slightly offset. The single tool failure was a multi edit that applied 15 of its 17 changes, so the most accurate model at tool calling. It was 3x slower than the other two, but it was the only one that ran an accessibility audit tool on its own page.

Live builds, screenshots, the full comparison table with the exact prompt I ran are here: [https://basalt.jgulfo.com/showcase/photographer/](https://basalt.jgulfo.com/showcase/photographer/)

Anyway, what are your thoughts? I think for web UI specifically 27B is more than enough for a very professional looking product, with a little bit of back and forth and more proper branding/UX design I can see any of them being a production website. I also can't stop being amazed at how a IQ1 quant can still deliver such high quality results even if the param count is multiple times over the two other models.  
  
\---

Disclosure: I'm the maintainer of the engine and harness all these models ran on.
```



document fetched 2026-11-18 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xyujv0/i_compared_qwen_38_27b_vs_38_fn_vs_ds_v41_flash
