{"slug": "moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push", "title": "Moonshot released Kimi K3's weights on July 27, extending its open-weight push", "summary": "Moonshot released the weights for its 2.8-trillion-parameter Kimi K3 model on July 27, 2026, 11 days after launching the model across Kimi, Kimi Work, Kimi Code, and the Kimi API, extending CEO Yang Zhilin's long-context research into an open-weight distribution strategy. The release marks a shift from Moonshot's earlier closed, consumer-first approach, allowing outside developers to deploy and extend the model, which features a one-million-token context window.", "body_md": "# Moonshot released Kimi K3's weights on July 27, extending its open-weight push\n\n**Moonshot launched the 2.8-trillion-parameter model on July 16 and released its weights 11 days later, pairing a one-million-token context window with CEO Yang Zhilin's long-running research thesis.**\n\nBy [RuntimeWire Staff](/author/runtimewire-staff)\n· Published\n\nPrimary source: [Aligned News - AI Intelligence](https://x.com/SonglinYang4/status/2078642849244205467)\n\n## Why it matters\n\nK3 carries Yang's long-context research into Moonshot's agent products while documenting a material strategy change. Moonshot moved from a closed, consumer-first position in 2023 to distributing model weights that outside developers can deploy and extend.\n\nMoonshot released [Kimi K3's weights](https://www.kimi.com/en/blog/kimi-k3?ref=runtimewire) on July 27, 2026, 11 days after launching the model across Kimi, Kimi Work, Kimi Code and the Kimi API. The release extends co-founder and CEO [Yang Zhilin's](https://kimiyoung.github.io/?ref=runtimewire) long-running focus on context and agentic systems into an open-weight distribution strategy. Kimi is available at [kimi.ai](https://kimi.ai/?ref=runtimewire).\n\nA [2023 GeekPark interview](https://about.geekpark.net/different-organizations-allow-different-things-to-emerge-yang-zhilin-s-first-in-depth-interview/?ref=runtimewire), conducted shortly after Moonshot introduced the original Kimi assistant, set out Yang's case for long context and closed consumer models. A later podcast conversation returned to K2 and agentic LLMs. Together, the interviews trace the technical argument behind K3 and the distribution strategy Moonshot changed along the way.\n\n[Watch: Zhang Xiaojun's conversation with Yang Zhilin on K2 and agentic LLMs](https://www.youtube.com/watch?v=ouG6jrkECrc&ref=runtimewire)\n\nYang earned his computer-science PhD from Carnegie Mellon University in 2019 after studying at Tsinghua University. His [biography](https://kimiyoung.github.io/?ref=runtimewire) lists research experience at Google Brain and Meta AI and his work on Transformer-XL and XLNet. Long context was already Yang's research problem before it became Kimi's product pitch.\n\n### The thesis arrived before the product\n\nIn the 2023 [GeekPark interview](https://about.geekpark.net/different-organizations-allow-different-things-to-emerge-yang-zhilin-s-first-in-depth-interview/?ref=runtimewire), Yang described the Transformer as a new kind of computer. Parameter count served as its processor, in his analogy, while context length served as memory. Larger memory would expand the applications the system could handle, from analyzing documents to maintaining relationships across years of interaction.\n\nThat argument shaped the original Kimi assistant. It focused on long conversations and document-heavy knowledge work, supporting roughly 200,000 Chinese characters in a session, according to [GeekPark's interview](https://about.geekpark.net/different-organizations-allow-different-things-to-emerge-yang-zhilin-s-first-in-depth-interview/?ref=runtimewire). Yang saw that capability as an early step toward systems that could retain enough history to become useful over long periods, rather than treating every interaction as a fresh start.\n\nYang also framed frontier AI as an organizational problem. In the 2023 conversation, he argued that OpenAI had assembled an organization where established research, large computing clusters and internet-scale data could produce a coherent product. His account treated organizational design as a prerequisite for sustained work toward increasingly capable systems.\n\n[Moonshot says](https://www.moonshot.ai/about?ref=runtimewire) Kimi has tens of millions of professional users each month.\n\n### From memory to hands\n\nAccording to Moonshot's [K3 technical report](https://arxiv.org/abs/2607.24653?ref=runtimewire), K3 is a 2.8-trillion-parameter mixture-of-experts model with 104 billion activated parameters, native vision and a context window of up to one million tokens. Moonshot's technical materials describe a routed-expert design for the model. The model carries the memory thesis into an agent stack: a large context window can hold repositories, documents and prior tool results while the model works through longer tasks.\n\n### Moonshot shifted its model distribution strategy\n\nIn 2023, Yang argued that a consumer super-app would depend on closed models because ownership of the underlying system allowed a developer to preserve product differentiation. He described closed development as the route to a dominant consumer interface. ([GeekPark](https://about.geekpark.net/different-organizations-allow-different-things-to-emerge-yang-zhilin-s-first-in-depth-interview/?ref=runtimewire))\n\nMoonshot later said it was [open-sourcing Kimi-K2-Base and Kimi-K2-Instruct](https://kimi.ai/blog/kimi-k2?ref=runtimewire), then published K3's weights through its official model release. That distribution gives outside developers and infrastructure providers a reason to optimize, deploy and build around Kimi.\n\nThe distribution shift is practical. Moonshot competes with better-capitalized U.S. developers and a crowded group of Chinese model builders. Distributing weights shifts some deployment and integration work to outside developers, while API pricing gives Moonshot a path to monetize usage. It also invites scrutiny. In February 2026, Anthropic named Moonshot among Chinese developers it accused of using distillation campaigns against its models. The allegation has not been adjudicated. ([Anthropic](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks?ref=runtimewire))\n\nCapital gives Yang room to pursue both the model and the distribution push. [TechCrunch reported](https://techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets/?ref=runtimewire) on May 7 that Moonshot raised about $2 billion at a $20 billion valuation in a round led by Meituan's investment arm.\n\nK3's scale does not settle the capability race. A [joint preliminary assessment](https://www.nist.gov/news-events/news/2026/07/uk-aisi-caisi-preliminary-assessment-kimi-k3s-cyber-capabilities?ref=runtimewire) by the U.S. Center for AI Standards and Innovation and the UK AI Security Institute tested a small set of public and private cyber benchmarks. K3 performed significantly below leading U.S. cyber-capable models, reaching step 17 on average in a 32-step simulated network attack where the leading U.S. models reached 28.5. It outperformed [GLM-5.2](/models/z-ai/glm-5.2:batch) in the same preliminary cyber evaluations, while its safeguards allowed attempts at exploit development and offensive operations. The agencies ran a selective set of cyber tests, so the results do not establish K3's broader coding or knowledge-work performance.\n\nYang's throughline is the choice of problems: context as memory, tools that let models act and model development as an organizational discipline. His distribution plan proved negotiable. K3 packages that long-running technical thesis into the form Moonshot currently needs: a downloadable model that can attract developers globally while creating a path to Moonshot's assistant, coding products and API.", "url": "https://wpnews.pro/news/moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push", "canonical_source": "https://runtimewire.com/article/kimi-k3-yang-zhilin-long-context-open-weight-agents", "published_at": "2026-08-21 16:22:30+00:00", "updated_at": "2026-08-21 16:45:36.604426+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-research"], "entities": ["Moonshot", "Kimi K3", "Yang Zhilin", "Kimi", "Kimi Work", "Kimi Code", "Kimi API", "Carnegie Mellon University"], "alternates": {"html": "https://wpnews.pro/news/moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push", "markdown": "https://wpnews.pro/news/moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push.md", "text": "https://wpnews.pro/news/moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push.txt", "jsonld": "https://wpnews.pro/news/moonshot-released-kimi-k3-s-weights-on-july-27-extending-its-open-weight-push.jsonld"}}