The number that got my attention isn’t the 2.8 trillion parameters. It’s that Kimi K3 won three of six agent tests, and on Artificial Analysis it beat Fable 5 outright on automation tasks. Agentic work is most of what I actually do with a model — run a tool, read the output, decide the next step. That’s the column I read first, and it’s the one where an open-weights model just landed in front.
So the practical question isn’t “is it as smart as Fable 5” (three Intelligence Index points say not quite). It’s whether the gap still justifies a closed API for the parts of my stack that are mostly loops and tool calls. Open weights means I get to answer that with my own eval set instead of trusting anyone’s internal benchmarks. That’s the whole point of weights you can download.
One caveat I’d hold onto: the hallucination rate climbed noticeably over K2’s 51 percent. Fable 5 at 55 and GPT-5.6 Sol at 85 are worse, which tells you something uncomfortable about the whole field rather than something reassuring about Kimi. For an agent with real permissions, a model that invents an answer instead of admitting it doesn’t know is the failure mode that costs you.
The story — Moonshot AI, backed by Alibaba, released Kimi K3, a 2.8-trillion-parameter open-weights model that its own benchmarks place just behind Claude Fable 5 and GPT 5.6 Sol, winning two of six coding disciplines and three of six agent tests. Independent lab Artificial Analysis scored it 57 on its Intelligence Index — a jump from rank 17 to rank 3 — with particularly strong agentic and automation results (Source).