Skip navigation
Hy-4 versus Deepseek Flash.
0:00 / 0:30

Hy-4 versus Deepseek Flash.

Loktar 🇺🇸

@loktar008.2K followers

Visit on X
212 likesDownloadOriginal post

11,200 views

Hy-4 versus Deepseek Flash. Using Deepseek Harness. I still really love the Deepseek output but Hy4 put up a great showing with about a quarter of the tokens and a 3rd of the requests.

28 Comments

Comments from the original post on X (Twitter)

  • Neo's profile picture

    @NeoAIForecast

    You are a master of these tests! This one is cool

    4 likes

  • Loktar 🇺🇸's profile picture

    @loktar00

    🙏 Thank you man I appreciate it!

  • Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)'s profile picture

    @teortaxesTex

    how do you make comparison videos?

    3 likes

  • Loktar 🇺🇸's profile picture

    @loktar00

    Wrote an app a long time ago that I modify for ffmpeg, you can def vibecode something like this super easy with Opus, it's great with videos!

    4 likes

  • Chromadera's profile picture

    @Chromadera

    Yeah. Hy4 is quite close. Seems it wasn't Benchmaxxed. If they release a flash that is good and we can run locally who knows.

    2 likes

  • Tyler Folkman's profile picture

    @tfolkman

    I wonder how large the weights will be.

    1 like

  • Loktar 🇺🇸's profile picture

    @loktar00

    770B with 49b active -

  • Tyler Folkman's profile picture

    @tfolkman

    Thanks!

    1 like

  • FancyRocket's profile picture

    @ldreisi

    Will there be a HY 4 Flash? Like 120B

    1 like

  • hlitoroy's profile picture

    @hlitoroy

    omp or nothing

    1 like

  • yifeng weng's profile picture

    @GeekerStack

    When using the same harness and having different models work on the same task, the second model may copy the first model’s answer and make adjustments. What I want to confirm is whether your comparison fully isolates memory, context, and the file system independently—only then can it compare their true capabilities.

    1 like

  • Loktar 🇺🇸's profile picture

    @loktar00

    I feel like this is an AI reply however they ran like a month apart in different containers so 99% sure that's impossible.

    1 like

  • yifeng weng's profile picture

    @GeekerStack

    You called it—that’s exactly what I said, haha: 使用同一个harness,用不同模型做同一个题目时,第二个模型会抄第一个模型的答案并做调整,想确认的是,你的对比是否有把记忆、上下文、以及文件系统做完全的独立隔离,这样才能对比出真正的水平

    1 like

  • s.h's profile picture

    @hebbarmp

    Great optimization from Hy4. 👏👏

    1 like

  • MabraX's profile picture

    @MabraX

    Sweet, deepseek is a beast still!

    1 like

  • DEXTUNE's profile picture

    @dextune

    Both have great quality. I think it's because your prompt was so well-made that we got such good results. Awesome!

    1 like

  • J A Z I I's profile picture

    @notjazii

    omg, wasn't expecting this better from hy4

    1 like

  • Chris W's profile picture

    @Chris_Wozniczek

    wow, deepseek looks a bit more detailed but Hy is a super positive surprise here also looking the token count. wow.

    1 like

  • Sal's profile picture

    @S4lXLV

    Could you please tell me what you use fro these? Are these just threejs or blender? Or something else

    1 like

  • Loktar 🇺🇸's profile picture

    @loktar00

    Just prompting and telling it to use three.js

    2 likes