reuters.com
www.reuters.com

Summary

Alibaba has launched Qwen 2.5-Max, an AI model it claims outperforms DeepSeek-V3, OpenAI’s GPT-4o, and Meta’s Llama-3.1-405B.

The release, coinciding with Lunar New Year, reflects mounting competition in China’s AI sector after DeepSeek’s rapid rise.

DeepSeek’s recent advancements have pressured Chinese rivals like ByteDance and Baidu to upgrade their models and cut prices.

DeepSeek’s founder downplays price wars, focusing on artificial general intelligence (AGI). The company’s lean, research-focused structure contrasts with China’s tech giants, which face challenges in AI innovation.

    • werefreeatlast@lemmy.worldEnglish
      1·
      7 hours ago

      Spitroastgroup 2.7 is coming out too! On no wait, its going back in. Nope, out. Ah actually in. Man they’re just going in and out like mad.

  • Breve@pawb.socialEnglish
    70·
    2 days ago

    Well, the models start comin’ and they don’t stop comin’…

    The US tech sector has just been completely disrupted. Turns out decades of slashing public education and demonizing “liberal” colleges is starting to catch up. Even Elmo himself said that H1B visas are critical because the US simply isn’t producing enough talent, but he and the other tech billionaires didn’t realize that money can’t buy everything, as they are now being shown with their pants down.

  • Em Adespoton@lemmy.caEnglish
    434·
    2 days ago

    DeepSeek’s “big change” isn’t the performance of its model though; it’s that it is fully open and operates on a fraction of the resources.

    Is alibaba’s model also open weights, open reasoning, free for anyone to run, and runnable (and trainable) on consumer hardware?

    • trevor@lemmy.blahaj.zoneEnglish
      401·
      2 days ago

      Call it “open weight” if you want, but it’s not “fully open”. The training data is still proprietary, and the model can’t be accurately reproduced. It’s proprietary in the same way that llama is proprietary.

      • Gsus4@mander.xyzEnglish
        112·
        2 days ago

        But I could use it as a starting point for training and build from it with my own data. I could fork it. I couldn’t fork llama, I don’t have the weights.

        • trevor@lemmy.blahaj.zoneEnglish
          111·
          2 days ago

          You can also fork proprietary code that is source available (depending on the specific terms of that particular proprietary license), but that doesn’t make it open source.

          Fair point about llama not having open weights though. So it’s not as proprietary as llama. It still shouldn’t be called open source if the training data that it needs to function is proprietary.

  • r00ty@kbin.life
    312·
    2 days ago

    Oh, good. Maybe they will stop trying to scrape my websites at some ridiculous rate using faked real browser UAs. I just blocked their whole ASN (AS45102) in the end.

    • ThePowerOfGeek@lemmy.worldEnglish
      21·
      2 days ago

      I already have the Temu AI psuedocode. Here you go:

      10 print “Hi, how can I help?”

      20 receive input

      30 print “That’s awesome! What else?”

      40 go to 20

  • Hubi@feddit.orgEnglish
    15·
    2 days ago

    Any word on the training cost? I feel like that’s the most relevant metric here.

  • Bronzebeard@lemm.eeEnglish
    61·
    2 days ago

    Oh cool, I was worried my 401k had almost sort of recovered from the last bombshell earlier this week…

  • A_A@lemmy.worldEnglish
    38·
    2 days ago

    DeepSeek_R1 outperform or equalzz GPT-1o is major newZ, but : 4o is much better than 1o. Now, Qwen-2.5Max outperforms GPT-4o … watever the investment involved, this is even more important ( ! ).