跳过导航
Ling-3.1-flash is now open source: ~560B total parameters, ~25B active per token and up to 1M…
0:00 / 1:42

Ling-3.1-flash is now open source: ~560B total parameters, ~25B active per token and up to 1M…

Ant Ling

@AntLingAGI1.6万 位关注者

在 X 上查看
1,991 个赞下载原帖

1,239,114 次观看

Ling-3.1-flash is now open source: ~560B total parameters, ~25B active per token and up to 1M context. Built for real work, it can turn a rough brief into a page, mobile UI or interactive prototype teams can review and refine. See the demo.

23 条评论

评论来自 X (Twitter) 上的原帖

  • Ant Ling 的头像

    @AntLingAGI

    For local serving, MXFP4 experts and block-FP8 dense layers cut weights from ~534GB to ~292GB; a tested Hopper setup moved from 8 GPUs to 4. In separate tests, 8 GPUs averaged 604ms to first token and 4 GPUs reached ~1.36 QPS. Workloads differed.

    3 个赞

  • Ant Ling 的头像

    @AntLingAGI

    For a business assistant, capability matters alongside response speed. On the Artificial Analysis leaderboard, Ling-3.1-flash pairs an Intelligence Index of 41 with ~210 output tokens/s. This helps teams assess both dimensions for frequent use.

    3 个赞

  • Ant Ling 的头像

    @AntLingAGI

    High-frequency use also has a budget constraint. A second Artificial Analysis view plots output speed against blended price: Ling-3.1-flash sits at ~210 tokens/s and ~$0.19 per 1M tokens. Provider prices may differ from this benchmark measure.

    2 个赞

  • Ant Ling 的头像

    @AntLingAGI

    What can teams build? Ling-3.1-flash drafts React Native screens, editable SVGs, websites and Web 3D prototypes. In Design Arena Open Weights, it ranks #3 in React Native and 3D, #4 in SVG and #5 in Website. Review the outputs in your own workflow.

    3 个赞

  • Ant Ling 的头像

    @AntLingAGI

    Thanks to @sgl_project and @vllm_project for supporting the Ling family across open inference stacks. Their work gives developers more ways to deploy and tune Ling models in their own infrastructure.

    2 个赞

  • Ant Ling 的头像

    @AntLingAGI

    Get the open weights and model docs: Hugging Face: ModelScope:

    4 个赞

  • Ant Ling 的头像

    @AntLingAGI

    Join the Ant Ling community on Discord:

    3 个赞

  • AJ 的头像

    @ItsmeAjayKV

    Finally, been waiting for this, as someone who used Ling-3.1-flash extensively via api i can tell you have done a phenomenal job with it.

    1 个赞

  • Chasen 的头像

    @chasen_liao

    开源了!Ling 这个模型越用越好用(òωó)👍

    1 个赞

  • Passenger0522 的头像

    @Passenger0522

    太期待了 终于等到你!

    1 个赞

  • 伟大 的头像

    @Huouo908070

    哇塞,开源了,之前使用过,效果非常好,模糊的需求就可以生成质量很高的作品。

  • 冰零 的头像

    @cryozerolabs

    终于开源了!可以升级到 3.1 了!

  • Lonely 的头像

    @Lonely__MH

    终于等到你🎉

  • Phong Designs AI Systems 的头像

    @phongdesigns

    review and refine is the part i'd want demoed. a generated page always looks right on its own. i had two copies of the same screen this week, one in the brand font and one not, and both passed a look-at-it review

  • atomic.chat 的头像

    @atomic_chat_hq

    congrats on the open-weights drop 🔥 already cooking quants for it :)

  • 程序员端哥 的头像

    @duange6099

    总参数约 560B,每 token 激活约 25B 量化后权重从约 534GB 压到约 292GB(MXFP4 专家 + block-FP8 稠密层) 实测 Hopper:8 卡收到 4 卡;8 卡首 token 约 604ms,4 卡约 1.36 QPS

  • s.h 的头像

    @hebbarmp

    🔥🔥

  • Modengsir AI 的头像

    @ModengSir

    等了好久 终于等到了,可以好好体验下大模型的能力了

  • 马识途 的头像

    @shitunote

    I've tested it. It is pretty cool .

  • Reid Vale 的头像

    @reid_vale

    I tested it myself, and this model has really high aesthetic standards - it actually looks even better than the frontend design I spent months working on, hahaha