China’s Kimi Raises $500M— While a“Post-Code”AI Quietly Reaches $250M ARR
Kimi's Overseas API Revenue Grows 4×
In the final days of 2025, two Chinese AI delivered unexpectedly positive signals to the market.
Shortly after Meta acquired Manus at a premium valuation, Moonshot AI—better known for its flagship model Kimi—announced a $500 million Series C round, valuing the company at $4.3 billion post-money.
The round was led by IDG Capital, heavily oversubscribed by existing shareholders, including Alibaba, Tencent, Gaorong Capital, and Wang Huiwen.
In an internal memo, founder and CEO Yang Zhilin revealed that the company now holds more than RMB 10 billion (over $1.4 billion) in cash on its balance sheet.
What stands out most is not the valuation, but the pace of commercialization. According to Yang, between September and November 2025, Kimi’s paid user base—across both domestic and international markets—grew at a monthly rate exceeding 170%. Over the same period, overseas API revenue increased fourfold.
This acceleration is closely tied to Kimi’s technical breakthroughs over the past year, particularly the release of K2 Thinking. Strong product progress drove revenue momentum, which in turn made this financing round one of the most competitive deals in China’s AI market.
From Large Models to Thinking Agents
K2 and K2 Thinking, positioned respectively as a large-scale foundation model and a reinforced reasoning model, represent a real step forward in complex reasoning and long-chain thinking. Kimi not only released China’s first trillion-parameter-scale model, but also open-sourced the country’s first truly agentic thinking model—achieving benchmark performance that matches or even exceeds comparable OpenAI models across several core evaluations.
K2 Thinking can be described as a genuine thinking model—one capable of hundreds of autonomous tool calls. The key shift is that intelligence no longer resides in a single static model, but in a reasoning agent that can continuously reflect, verify, and invoke tools while solving complex tasks.
In practice, the model can autonomously execute 200–300 rounds of tool usage without human intervention. This marks a transition from passive response generation to active problem solving.
Choosing Private Capital Over IPOs
While peers such as Zhipu and MiniMax are accelerating toward public listings, Yang has taken a different view. He argues that private markets still offer deeper pools of capital. In fact, Kimi’s combined Series B and C raised more capital than most IPOs or follow-on offerings by public companies. As a result, the company feels no pressure to go public in the near term—and is not optimizing for an IPO.
In my opinion, Anthropic carved out a distinct path by focusing on breakthroughs in coding and agent capabilities, rather than following OpenAI’s roadmap directly.
For Chinese foundation-model startups—often operating under tighter capital constraints—this realization is spreading quickly. OpenAI’s path is not the only path. More teams are now explicitly aligning themselves with Anthropic’s direction, and Kimi began investing heavily in agentic capabilities as early as the K2 generation.
K3: Scaling FLOPs and Product Intelligence
Looking ahead, Yang has made his ambitions explicit: Anthropic is the benchmark, and the goal is to surpass frontier peers to build a world-leading AGI company. That ambition is crystallized in K3, whose roadmap combines architectural improvements with aggressive scaling.
The plan is to increase effective FLOPs by at least an order of magnitude—pushing Kimi’s pretraining compute closer to the global frontier. In practice, this means K3 is not merely a larger model, but one that is more deeply trained, more compute-efficient, and more capable of generalization.
Just as important is tighter vertical integration between model training and agent product “taste.” The goal is for K3 to lead not only on benchmarks, but in real-world usage—delivering differentiated agent behavior and user experience rather than abstract score gains.
In a recent Reddit AMA, the team hinted at potential architectural shifts, including KDA (Kimi Delta Attention) or similar mechanisms aimed at overcoming Transformers’ limitations in long-context reasoning and efficiency. By exploring linear or optimized attention, Kimi is signaling a move away from brute-force scaling toward structural intelligence—a transition that could prove decisive for the next generation of frontier models.
The Rise of “Post-Code” AI and a $250M ARR AI DevOps Giant
Earlier this week, there was a hot talk about Background Agent Infrastructure by Rahul, the head of Applied AI at Ramp, and its role in pushing software systems toward greater autonomy.
Coincidentally, one product closely aligned with this trend has already reached $250 million in ARR, growing at 50% year over year.
