• qaz@lemmy.world
    link
    fedilink
    English
    arrow-up
    7
    ·
    edit-2
    8 hours ago

    According to artificialanalysis.ai’s latest benchmarks, it scores better than Opus 4.8 set to max, despite costing half as much per task.

    It also beats the top models of some of the largest US tech companies such as xAi, Meta, Google, and Nvidia.

    I wonder what the US tech investors will think of this, and what this will mean for the financial AI bubble.

  • hummingbird@lemmy.world
    link
    fedilink
    English
    arrow-up
    3
    ·
    7 hours ago

    Despite being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.

    So they admit that the benchmarks are bulshit? Lol.

  • mal3oon@lemmy.world
    link
    fedilink
    English
    arrow-up
    10
    ·
    10 hours ago

    It’s not released as open-weight yet. They plan to do so on the 27/07, so I just hope that is not a tactic to build momentum then change their mind à la Elon for the hype. Nonetheless, performance wise it’s amazing, only downside is the size … 3T. Good luck running that on “consumer hardware”.

  • FaceDeer@fedia.io
    link
    fedilink
    arrow-up
    13
    ·
    11 hours ago

    Looks like China’s gone from being 6 months “behind” to more like 6 weeks “behind.”

    Kind of pricey, but still the cheapest in its weight class. Could be they’re simply squeezing what profits they can out of it. I read that they’re planning to open the weights in August, at which point third parties will be able to compete based on whatever the actual cost of running it is.

    • blueworld@piefed.world
      link
      fedilink
      English
      arrow-up
      2
      ·
      7 hours ago

      Sonnet 4.6 price equivalent, with half to a third of it’s tok/s. It’ll be interesting to see if the near gpt 5.6 or opus 4.8 level is accurate.

  • eicker@lemmy.world
    link
    fedilink
    English
    arrow-up
    9
    ·
    11 hours ago

    Open models crossing the 3T barrier this quickly is wild. 🤯 The interesting part isn’t the parameter count though: it’s whether developers can actually run, fine tune, and build useful products around it. If the benchmarks hold up in the real world, proprietary labs just got another serious competitor.

  • dgdft@lemmy.world
    link
    fedilink
    English
    arrow-up
    8
    ·
    14 hours ago

    I’m happy to see the Chinese labs start competing on SOTA… but hot damn that price is a turnoff.

    Anyone tried Kimi 2.7 Coding out in practice yet? Curious to know how it fares against Deepseek 4 and similar budget options.

    • very_well_lost@lemmy.world
      link
      fedilink
      English
      arrow-up
      10
      ·
      13 hours ago

      Looks like they intentionally priced it just a little less than Opus… probably to give the impression that it’s a cheaper alternative without people thinking it’s too cheap for a SOTA model and getting suspicious.

      As usual, the AI marketing is more sophisticated than the actual technology.

  • pikl@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    arrow-down
    6
    ·
    7 hours ago

    The world’s greatest thieves stole more than anyone else and compiled it into the world’s greatest bullshit generator… Impressive.