one year on
Moonshot releases Kimi K2: a trillion-parameter open-weights agent
China's second 'DeepSeek moment' of the year — a 1T-parameter mixture-of-experts model, open weights under a modified MIT license, tuned for agentic tool use rather than chat.
Moonshot AI releases Kimi K2, a 1-trillion-parameter mixture-of-experts model (32B active per token) — open weights, modified-MIT licensed, and explicitly built for agentic workloads: tool use, multi-step coding, autonomous task execution.
The benchmark story is immediate: state-of-the-art among open models on coding and agentic evals, competitive with closed frontier models on several. The economic story is louder: API pricing at roughly a tenth of Western frontier rates, and the weights are free.
The release lands the same day OpenAI confirms its own open-weights model needs “additional safety testing” — a delay announced as China’s labs ship weekly. Six months after DeepSeek R1 rearranged assumptions about where open AI leadership lives, K2 suggests that wasn’t an anomaly. It was a cadence.
The record
Within the weekend, K2 tops open-model coding leaderboards and the comparisons to Claude Sonnet — at a fraction of the API price — write themselves.
The uncomfortable observation: the most capable open-weights model in the world is, again, Chinese — and OpenAI's own long-promised open model is, again, delayed.
One year later — open only if you can handle spoilers
K2 held the open-weights crown until it was superseded by an even faster cadence of Chinese releases — and its successor K2 Thinking, in November 2025, briefly matched closed frontier models on reasoning benchmarks. The 'open weights = one year behind' assumption did not survive 2025.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy