Moonshot's Kimi K3: A 2.8-Trillion-Parameter Open Model — Strong, but Behind the Leaders

Open-source AI has edged a little closer to the frontier models. China's Moonshot released its new Kimi K3 and calls it "the world's first open 3T-class model" at 2.8 trillion parameters. It's especially strong at long-horizon coding and autonomous agent work. What stands out is Moonshot's candor: its own announcement says the model trails Claude Fable 5 and GPT-5.6 Sol and has a noticeable gap in user experience.

In brief

• What launched: Kimi K3, Moonshot's open-source model. 2.8 trillion parameters (an MoE with 896 experts, 16 active), a 1-million-token context window, and native vision. Full weights are due by July 27.
• What it's good at: long-horizon coding (chip design, GPU compilers), game development, and research automation. It says it completed an astrophysics research task in about two hours that would normally take one to two weeks.
• Candid limits: Moonshot's own announcement lists shortcomings — the model trails Claude Fable 5 and GPT-5.6 Sol, shows "excessive proactiveness" (making autonomous decisions mid-task), and has a noticeable gap in user experience.
• Price + access: on the API, cache-hit input is $0.30, cache-miss input $3, output $15 (per 1M tokens). Available today via Kimi.com, Kimi Work, Kimi Code, and the API.

Our take

This isn't a "use it tomorrow" tool for you: running a 2.8-trillion-parameter model on your own machine isn't practical for most of us, and the API price swings a lot depending on the task. But two things matter. First, the direction: open-source frontier models keep growing and getting stronger, which means more choice and price pressure over time. Second, and more interesting, is Moonshot's stance — it plainly wrote that "we trail the leaders, our UX has gaps." In an industry where everyone declares their own model "the best," that candor is valuable in itself.

Kaynak: Moonshot Kimi