This Free Chinese AI Model Has Now Surpassed GPT-5, Claude Sonnet 4.5, and Grok-4

This week, a Chinese startup called Moonshot AI released a new model Kimi K2 Thinking and quietly flipped the leaderboard. In independent benchmarks that test reasoning, coding, and autonomous tool use, this free, open-source model didn’t just compete with OpenAI’s GPT-5.
It beat it.

Kimi K2 Thinking didn’t appear out of nowhere. It follows a fast-moving wave of Chinese open-weight models like DeepSeek R1, Qwen, GLM, and MiniMax-M2, which had briefly claimed the open-source crown just weeks ago.

And that changes the conversation around who really controls the future of AI.

From “Open Can’t Compete” to “Open Just Won”

Kimi K2 Thinking isn’t a lightweight experiment or a research demo. It’s a full-scale frontier model built to reason, plan, code, browse the web, and operate tools on its own, the kind of work typically reserved for premium, closed systems.

In third-party benchmarks widely used by AI researchers, K2 Thinking topped both proprietary and open competitors: It scored 44.9% on Humanity’s Last Exam, one of the toughest reasoning tests available, 60.2% on BrowseComp, which measures how well AI agents search and reason online, 71.3% on SWE-Bench Verified, a gold-standard test for real-world coding and Strong results across math, retrieval, and agentic workflows

In multiple categories, those scores surpassed GPT-5, Claude Sonnet 4.5 (Thinking), and Grok-4.

The headline isn’t just performance. It’s how the performance was achieved.

Open Source — With a Twist

Moonshot released K2 Thinking under a Modified MIT License, making it one of the most permissively licensed frontier models available today.

Developers can use it commercially, modify it, build products on top of it and deploy it on their own infrastructure

The Bigger Shift No One Can Ignore

This moment isn’t just about benchmarks. It’s about power.

The most capable reasoning model available today isn’t locked behind a paywall. It’s open, inspectable, adaptable, and free to use. 

Instead of asking “How big can models get?”, the industry is starting to ask a different question:

Who can build the smartest systems sustainably?

Kimi K2 Thinking suggests the answer may come from efficiency, openness, and smarter architecture, not just deeper pockets.

And that may be the most disruptive idea in AI right now.