aivs.tv

AI Coding

Voxel world on one Mac Studio, local Flash models — Qwen 3.8 Flash Next vs DeepSeek V4.1 Flash 1017

0:00 / 0:38
Qwen 3.8 Flash Next (top)
Posted byNiklas@niklaslenz_ai · on X
View post on X

4 likes407 views

Scoreboard2

  1. Qwen 3.8 Flash NextAlibaba (Qwen)
    Time
    58.1 min
    Tokens
    225.9K
    • Top; oQ4e quant with MTP, run in Qwen Code 0.25.2; 58.1 agent minutes, ended by the tool-call cap
  2. DeepSeek V4.1 Flash 1017DeepSeekDeepSeek Harness
    Author’s pick
    Time
    1 h 06 min
    Tokens
    168.2K
    • Bottom; LSQ3-G128 quant, DeepSeek Harness 0.2.1-rc.2; 65.7 agent minutes

Prompt

Voxel world on one Mac Studio, local Flash modelsThe post doesn’t quote the prompt wording.

Author’s verdict
The author likes the DeepSeek world much better, even though Qwen decodes about 2.4 times faster.

@niklaslenz_ai’s view (paraphrased)

Qwen 3.8 Flash Next (top) and DeepSeek V4.1 Flash 1017 (bottom) run locally on one 256 GB Mac Studio with the same voxel-world prompt, each in its maker’s own coding agent, thinking on, one try.

The post, itemised

Kind
Side-by-side
Task
Voxel world on one Mac Studio, local Flash models
Category
3D & Shaders
Mode
Agentic
Posted
Video length
38 s
Likes
4
Views
407
Bookmarks
2
Sponsored
No

NotesBoth models ran quantized on oMLX on the author’s own machine, so results reflect these local quants, not the hosted APIs.

Cost & time, as stated in the postAs stated: Qwen 3.8 Flash Next 58.1 agent minutes, 225,892 tokens, browser check 8.5/10; DeepSeek V4.1 Flash 1017 65.7 minutes, 168,228 tokens, 8/10.

Related

Qwen 3.8 Flash NextDeepSeek V4.1 Flash 1017DeepSeek V4.1 Flash 1017 vs Qwen 3.8 Flash Next

Posted by @niklaslenz_ai on X. aivs.tv claims no ownership: it hosts a copy for fast playback and links to the post on X. Request removal

More with these models