Loading...
Loading...
Moonshot's Kimi K3 packs 2.8 trillion parameters and a million token context, beats several frontier models on reasoning benchmarks, and the open weights land next week. Here's what happened when I tried it.
The Chinese labs keep shipping, and Kimi K3 might be the strongest argument yet that open models belong in the same sentence as frontier ones.
Across coding and agentic benchmarks it edged past Opus 4.8 and GPT-5.5, trailing only the very top tier. API access was available immediately, so I ran my usual gauntlet: summarize a long spec, refactor a small module, explain a regex I wrote at 2am and no longer understand.
The long context test. I fed it an entire design doc plus three weeks of chat history from our team channel, roughly 400k tokens, and asked for a decision log. It produced something I would have paid a contractor for.
Previous attempts with smaller context models turned into a game of telephone where details mutated at every hop.
The 2am regex deserves its own paragraph. It parsed timestamp ranges from our alert format and contained a nested lookahead I no longer understood sober. Every model I'd asked before either rewrote it into something subtly wrong or explained it incorrectly. K3 traced the logic line by line, spotted that one group captured greedily, and suggested the exact character class fix. Then it wrote three test cases that would have caught the original bug.
Reports pegged complex task costs around ten dollars per completed job when the agent takes eighty-plus turns. Still cheaper than a contractor, but not the pennies-per-task story the raw token price implies.
Still, watching an open model trade blows with the closed top tier stopped feeling theoretical this week. My TACOS chatbot backend is next in line to get a K3 based fallback route.
Cheers, Yassen