November 6, 2025
Trillion params, trillion opinions
Kimi release Kimi K2 Thinking, an open-source trillion-parameter reasoning model
Kimi drops a trillion-parameter brain; comments clash over China’s lead and size vs local
TLDR: Kimi launched a giant open-source “thinking” AI that chains hundreds of steps to ace tough tests. Comments split between cheering China’s open-source surge and demanding smaller, local models, with playful pelican demos and bold claims it rivals top closed systems.
Kimi just unleashed a gigantic open‑source “thinking” model that chains hundreds of steps, clicks around the web, runs code, and scores big on exams with scary names like Humanity’s Last Exam. The vibe? Big brain, bigger reactions. Fans cheered the agent’s brag—200–300 tool calls—and its benchmark wins, while casuals said it sounds like an AI that can plan a research project and then actually do it. You can try it at kimi.com or via the API, and a full “agent” mode is coming soon.
Then the comments lit up. One camp crowed that China is now the open‑source ringleader, pointing to a streak from DeepSeek, Qwen, Kimi, and GLM, while teasing the US and Europe for “ghosting” open releases. Another camp rolled its eyes at trillion‑anything, begging for smaller models you can run locally—“don’t need a supercomputer to fix a bug.” A bold voice declared K2 “better than every closed model except GPT‑5 Codex” (spicy, unverified), and someone spun up a pelican riding a bicycle with a one‑liner command, instantly becoming the meme mascot. Between patriotic chest‑thumping, size‑vs‑smarts debates, and DIY demos, the crowd turned benchmarks into popcorn fodder—is this the future of AI, or just the world’s smartest Rube Goldberg machine?
Key Points
- •Moonshot AI introduced Kimi K2 Thinking, an open-source thinking agent model.
- •The model performs tool-augmented reasoning during inference and can execute 200–300 sequential tool calls.
- •K2 Thinking reports state-of-the-art results: 44.9% on HLE (with tools), 60.2% on BrowseComp, and 71.3% on SWE-Bench Verified.
- •Evaluations also reference SWE-Multilingual and LiveCodeBench V6 for coding and competitive programming.
- •K2 Thinking is available on kimi.com (chat mode) and via the Kimi K2 Thinking API; full agentic mode is coming soon.