Moonshot AI Bets on Long Context, Hits $35B

What the business is

Moonshot AI is a Chinese large model startup that self-develops a 100-billion-parameter general large model, focuses on the consumer-side intelligent assistant Kimi, which supports ultra-long input, and later pivoted to open-source models and commercialization of Agent capabilities

Starting capital：After its 2023 founding, it raised nearly RMB 2 billion in early funding (Sequoia (红杉), Capital Today (今日资本), ZhenFund (真格基金), etc.), with a valuation of over $300 million

How it started

In March 2023, Yang Zhilin (first author of Transformer-XL and XLNet), Zhou Xinyu (one of the authors of ShuffleNet), and Wu Yuxin (author of detectron2), all from Tsinghua, founded Moonshot AI. The team judged that the “emergence” of large models relied on 100 billion parameters, but the deployment bottleneck had a second dimension—how much context could be processed per conversation round, i.e., the model's “memory”. In October 2023, the company released Kimi Chat for the consumer side, supporting 200,000-character input, setting the then-global record for productizable long text; Yang Zhilin said, “Long-text technology is the first step of large models' moonshot program.”

What happened

In 2023, cumulative funding was nearly RMB 2 billion (Sequoia, Capital Today, Lisi Capital (砺思资本), etc.), with a valuation of over $300 million. In February 2024, it completed a round of over $1 billion; investors included Sequoia China (红杉中国), Xiaohongshu (小红书), Meituan (美团), and Alibaba (阿里), with a post-money valuation of about $2.5 billion, the largest single-round funding for a domestic large model at the time. In 2024, Kimi conducted large-scale paid traffic buying; in October, advertising spend reached RMB 199 million, and monthly active users grew from about 4 million to 22 million in half a year. After the open-source shock of DeepSeek-R1 in 2025, the company shifted to model foundations and Agent: in July, it open-sourced the trillion-parameter-level MoE model Kimi K2; one week after release it topped open-source leaderboards (GitHub 5.6K stars, Hugging Face downloads nearly 100,000, Perplexity announced it would do post-training based on K2); the architecture inherited DeepSeek V3 and made four parameter adjustments. In December 2025, it completed a $500 million Series C (valuation $4.3 billion, cash on hand over RMB 10 billion); in April 2026, it open-sourced K2.6 and disclosed that ARR exceeded $200 million; in May, it completed a $2 billion Series D (led by Meituan Longzhu (美团龙珠), post-money valuation exceeded $20 billion); in July, Series F exceeded $3.5 billion, post-money valuation $35 billion; Series G (Pre-IPO) started early, pre-money valuation rose to $50 billion.

What can be applied

When rivals competed on compute, long context cracked the market; but open source flattened it, so advantage had to become model capability and commercial loop—funding pace determined survival.