Moonshot AI: Kimi Models, Open Weights & Company History

A Beijing lab with Alibaba behind it

Moonshot AI was founded in Beijing in early 2023 by Yang Zhilin along with Zhou Xinyu and Wu Yuxin, former schoolmates at Tsinghua University. The company describes its founding team as bringing together people who worked on several earlier AI systems, without naming which ones. Its consumer product is Kimi, an assistant built around long-context reasoning: holding onto a conversation, a document set, or a coding session well beyond what most chat products keep in memory. [1][2]

Alibaba led a reported $1 billion round in February 2024 that valued the company near $2.5 billion, and Tencent joined a further round later that year. Reporting in mid-2026 put Moonshot's valuation at roughly $35 billion, behind only DeepSeek among privately held Chinese AI companies. Moonshot itself does not publish funding figures on its own site; these numbers come from press reporting on the rounds, not from a Moonshot filing or announcement. [3][2]

Open weights as the strategy

Moonshot's relevance to the field runs through a choice most large Chinese labs have made since DeepSeek's R1: publish the weights rather than gate the model behind an API alone. A downloadable trillion-parameter model is usable by anyone with the hardware to run it, which turns a single company's training run into infrastructure other companies and researchers build on, for better or worse depending on who is asking. [4]

Atlas interpretation: That strategy also sets Moonshot apart from Western frontier labs, most of which keep their largest models closed. Racing to publish weights at a scale nobody else has shipped is itself a form of competition: it is a claim about engineering capability that a benchmark score alone does not make as concretely. [4][8]

Four releases in thirteen months

Moonshot's events on the timeline show a fast release pace. Kimi K2, published in July 2025, was a trillion-parameter mixture-of-experts model with 32 billion active parameters, pretrained on 15.5 trillion tokens, and Moonshot billed it as the first open-weight model at that total parameter count. Downloads reportedly doubled over its first weekend, a notable figure for a file of roughly a terabyte. [4][5]

In November 2025, Kimi K2 Thinking added extended reasoning and chained tool calls, and Moonshot reported it beating several closed frontier models on agentic benchmarks. The US government's AI standards center published an independent evaluation of the model five weeks later, which is the kind of outside check open weights make possible: nobody has to take Moonshot's benchmark numbers on faith. [6]

Kimi K2.5, shipped in January 2026, kept the trillion-parameter mixture-of-experts design but added multimodal input and a mode Moonshot calls Agent Swarm, coordinating up to a hundred sub-agents on a single task. Moonshot reported it topping the HLE-Full benchmark and beating GPT-5.2 on several others; those are the vendor's own reported results, not an independent replication. [7]

Kimi K3 followed in July 2026 at 2.8 trillion parameters with 104 billion active, described by outside coverage as the largest open-weight model released to date. Downloading the full weights means moving roughly 1.4 terabytes of data, which is its own kind of barrier to actually running the thing yourself. [8][9]