What's this about?
Chinese lab Moonshot AI released Kimi K3, a 2.8-trillion-parameter model — the largest ever announced with a promise to publish the full weights (due July 27). It has a 1-million-token context window and two new architecture tricks (Kimi Delta Attention and Attention Residuals). Moonshot is candid that it still trails the two strongest proprietary models, Claude Fable 5 and GPT-5.6 Sol, but claims it beats everything else — including Claude Opus 4.8 and GPT-5.5. It's usable today via Kimi's website and API; the open-weights file lands within two weeks.
What exactly is a "3T-class open model"?
A model's size is measured in parameters — the learned numbers that make up its "brain." 2.8 trillion (~3T, rounded up) makes K3 the biggest model anyone has promised to release openly; the previous open-size record was DeepSeek's 1.6T. "Open" means the weights file itself will be downloadable, so companies can run it on their own hardware instead of renting access through an API — like owning the master recording instead of paying for streaming. The catch: at this size, "your own hardware" means a serious GPU cluster, not a laptop.
What's the community saying?
- The pricing raised eyebrows — $3/$15 per million tokens is "extremely high for a Chinese open-weight model," basically proprietary-frontier pricing; several read it as the end of the subsidized era for Chinese AI APIs.
- Is it worth it vs cheaper open rivals? — repeated comparisons to GLM 5.2 at less than a third of the price; the counter is that Artificial Analysis puts K3's per-task cost ($0.94) at half of Opus 4.8's.
- Benchmark skepticism — the launch page self-reports wins, the technical report isn't out yet, and the tech blog hadn't even been updated; "wait for the weights and independent evals."
- What does this mean for US labs? — if an open model is genuinely just behind Fable 5 and Sol, the moat discussion (echoed by oran_ge: "what's Anthropic's value then?") starts all over again.