Moonshot AI

Moonshot AI (月之暗面) was founded in Beijing in March 2023 by Yang Zhilin, alongside co-founders Zhou Xinyu and Wu Yuxin. Yang holds a PhD in computer science from Carnegie Mellon University and had worked as a researcher at Meta AI and Google Brain before returning to China; he is a co-author of the Transformer-XL and XLNet architectures, two influential extensions to the original Transformer that improved long-context modeling — expertise that fed directly into Moonshot's subsequent focus on extreme context lengths. The company's flagship product is the "Kimi" family of large language models and the associated Kimi chat assistant, which built an early consumer following in China on the strength of a then-novel 200,000-character context window. Yang Zhilin retains a 51.83% controlling stake in the company's domestic entity, with the four founders together holding roughly 65.6%. On July 17, 2026, Moonshot released Kimi K3, a mixture-of-experts model with 2.8 trillion total parameters — activating only 16 of its 896 experts (roughly 1.8% of the network) per token — which the company and independent commentators described as the world's first openly available "3-trillion-class" model, and the largest open-weight model released to date. That is roughly 75% larger than DeepSeek's V4 Pro, previously the largest widely used open model. K3 introduces two new architectural components, Kimi Delta Attention (a hybrid linear-attention mechanism) and Attention Residuals, built around a native multimodal input pipeline and a 1-million-token context window, explicitly targeted at long-horizon coding, reasoning, and autonomous-agent workloads. On released benchmarks, K3 trails Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol in overall capability, but it beat both Claude Opus 4.8 and GPT-5.5 on coding and general-agent evaluations. Moonshot also open-sourced a compact compiler it built internally, MiniTriton, and reported that AI-generated kernels reached 517.8 TFLOPS on an Nvidia H200 — a 14.82x speedup over an optimized PyTorch baseline — in benchmarking that Moonshot's technical blog said also involved "a GPGPU from an alternative vendor" it declined to name. At its API launch, Kimi K3's weights were not yet public; Moonshot promised a full open-weight release "by July 27, 2026." The raw checkpoint, quantized to MXFP4/MXFP8 from quantization-aware training, runs to roughly 1.4 terabytes, and Moonshot recommends serving the model on "supernode" clusters of 64 or more accelerators linked by high-bandwidth interconnect — no single Nvidia H100, H200, or B300 can hold the full model, and expert-parallel routing across hundreds of experts collapses in throughput once it has to cross a slower network fabric. K3's release also became the center of an export-control controversy. On July 22, 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged on social media that Moonshot AI had "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models," and that the company had built "a sophisticated internal platform to conduct large scale distillation against U.S. models" — widely understood to reference Anthropic's models — that let it "quickly switch between multiple methods of access to avoid detection." Treasury Secretary Scott Bessent said sanctions were "on the table," though no specific US government action against Moonshot had been announced as of this writing. The allegation followed a May 31, 2026 BIS memo reaffirming that Blackwell-class chips, including the GB300, require a license for export to Chinese entities; the Thailand reference points to a suspected transshipment route through Southeast Asia rather than direct import into China, and Kratsios's claims remain allegations rather than adjudicated findings — several AI researchers publicly pushed back within a day, arguing that K3's benchmark gains were better explained by its architecture and training recipe than by distillation from a rival lab's model. Moonshot's valuation has risen in step with Kimi's technical releases: roughly $4.3 billion at the end of 2025, $10 billion in the spring of 2026 after a $700 million raise, and $20 billion in May 2026 after a $2 billion round led by Meituan's Long-Z Investments (with Tsinghua Capital, China Mobile, and CPE Yuanfeng also participating) — $3.9 billion raised across six months, reportedly the largest financing run in China's large-language-model sector. Following K3's launch, Moonshot began raising a further round at a $30 billion valuation and was reportedly preparing a final pre-IPO round targeting a pre-money valuation of up to $50 billion, ahead of a planned Hong Kong listing expected within roughly six months of the K3 launch. Backers across these rounds include Alibaba, Tencent, HongShan (formerly Sequoia China), ZhenFund, IDG Capital, and — through a state-backed private equity vehicle — China's National Council for Social Security Fund. Moonshot now competes directly with DeepSeek, Zhipu AI, and Alibaba's Qwen team for the leadership position among China's open-weight model developers, and K3's benchmark showing — especially on coding and agentic tasks — has strengthened its case heading into an IPO. But the company also carries meaningfully more geopolitical exposure than domestic peers who have relied on Huawei Ascend hardware: the Kratsios allegations, if substantiated, would put Moonshot's compute-sourcing practices under far closer scrutiny from the same export-control apparatus that has already constrained Cambricon, Zhipu, and Huawei. Whether K3's open-weight release lands on schedule, and whether US regulators take further action, will likely shape both Moonshot's IPO timeline and the broader argument over how effectively chip export controls actually stop Chinese labs from reaching the frontier.