Moonshot AI, one of China’s fastest‑rising AI startups, has released a new flagship version of its Kimi model that it says pushes “open frontier intelligence” into territory previously reserved for US labs such as OpenAI and Anthropic. The Kimi K3 system, a 2.8‑trillion‑parameter mixture‑of‑experts model with native vision and a one‑million‑token context window, is already live in consumer apps and APIs, with Moonshot promising to publish its full weights by late July in what would be the largest open‑weight model ever released.
What Moonshot AI released
Moonshot AI announced Kimi K3 on July 16–17, describing it as its new flagship “frontier” model and the world’s first open‑weight system in the three‑trillion‑parameter class. The company’s tech blog and external analyses say K3 has 2.8 trillion total parameters, implemented as a mixture‑of‑experts architecture in which only a small subset of “experts” are active at any given time.
K3 powers the Kimi chatbot on web and mobile, Kimi Work for workplace productivity, Kimi Code for coding agents, and an OpenAI‑compatible API at api.moonshot.ai. Moonshot has also committed to releasing the model’s full weights by July 27, a step that would give researchers and enterprises direct access to an open‑weight system at this scale for the first time.
K3 is positioned as a long‑horizon reasoning engine rather than just a conversational bot: Moonshot says it is designed to handle large software repositories, operate tools like terminals and browsers, and combine coding with visual and video feedback.
Architecture: 2.8T mixture‑of‑experts, Kimi Delta Attention
Under the hood, Kimi K3 combines enormous scale with architectural tweaks meant to keep it efficient.
Moonshot’s tech note and independent blogs say K3 uses a Stable Latent Mixture‑of‑Experts design with 896 experts, of which only 16 are activated per token. This allows the model to reach 2.8 trillion parameters on paper while limiting compute per step, a pattern now common across frontier models.
The model also debuts Kimi Delta Attention (KDA), a hybrid attention mechanism that adapts to very long sequences and moves away from traditional prefix caching. In combination with Attention Residuals, KDA is designed to make 1‑million‑token context practical, with Moonshot claiming up to 6.3× faster decoding at that length compared with standard attention approaches.
DigitalApplied and AI/TLDR summaries say K3 uses MXFP4 quantized weights and MXFP8 activations from the supervised‑fine‑tuning stage onward, improving memory efficiency and throughput without large accuracy losses. Moonshot itself claims about 2.5× “scaling efficiency” versus its previous K2 model.
Together, these design choices make K3 a test case for whether ultra‑large, long‑context models can be opened to the broader ecosystem without becoming prohibitively expensive to run.
Context window and modalities: 1M tokens, native vision, and video
One of K3’s headline features is its one‑million‑token context window.
Moonshot says K3 can accept up to 1,048,576 input tokens and produce up to the same number in completions, though default completions are capped much lower. This capacity allows the model to process entire codebases, long legal documents, or multi‑session chat histories in a single “conversation,” instead of requiring chunking or external retrieval.
The model is also natively multimodal. It can take text, images, and video as inputs, with images uploaded directly and videos provided via a files API. Moonshot says K3 can inspect its own visual outputs, such as user interfaces or game screens, and then revise them based on the analysis.
K3 streams outputs as separate “reasoning” and “content” deltas, allowing clients to display the model’s internal thinking trace in near real time while the final answer is still being generated. That mirrors features in newer OpenAI and Anthropic models and is meant to support transparency and debugging for complex tasks.
Performance: where K3 sits in the frontier hierarchy
On benchmarks, K3 is strong but not quite at the very top.
Moonshot’s blog and external testing summarized by DataCamp and AI/TLDR say K3 scores around 57 on the Artificial Analysis Intelligence Index, placing it third globally behind Anthropic’s Claude Fable 5 and OpenAI’s GPT‑5.6 Sol. CNBC reports that Moonshot itself acknowledges K3 still trails Fable 5 and GPT‑5.6 overall, but claims it closes much of the gap and beats many rivals on specific tasks.
Reddit and third‑party breakdowns note that at maximum reasoning effort K3 achieves scores in the high 80s on Terminal Bench 2.1 (a tools and command‑line benchmark) and mid‑70s on Program Bench for coding. AI/TLDR says K3 ties or outperforms some US models on certain coding and web‑development benchmarks, and that it holds its own on long‑context browsing tests.
China Daily and Straits Times reporting emphasize K3’s performance in long‑horizon coding and knowledge work, with Moonshot positioning it for enterprise tasks that involve sustained reasoning over large bodies of information. On some frontier evaluation suites, K3 still lags Anthropic’s Fable 5 and OpenAI’s top models, but its open‑weight status gives it a different kind of strategic edge: adaptability.
Availability and pricing
Kimi K3 is already exposed across Moonshot’s product family and a public API.
AI/TLDR and Moonshot’s documentation say the API model name is kimi-k3, served from api.moonshot.ai with an OpenAI‑compatible interface. The model is available via kimi.com, Kimi’s mobile apps on iOS, Android and HarmonyOS, and in Kimi Code tiers that offer 256K or 1M‑token contexts depending on subscription.
Moonshot uses a three‑tier pricing structure per million tokens: about $0.30 for cache‑hit input, $3.00 for cache‑miss input, and $15.00 for output, according to Times of AI. China Daily notes that the API charges 100 yuan (roughly $14.70) per million output tokens, significantly higher than domestic competitors such as DeepSeek and Xiaomi’s MiMo V2.5 Pro, and above GLM‑5.2 and Alibaba’s Qwen3.7‑Max.
That premium pricing marks a departure from the low‑cost strategies many Chinese AI providers have used and suggests Moonshot is betting enterprise buyers will pay for higher capability and open‑weight flexibility.
Strategic context: China’s open‑weight answer to US frontier models
Kimi K3’s release is also a political and economic signal.
CNBC and Straits Times coverage highlight Moonshot’s claim that K3 “closes the gap” with leading US systems and positions China as a contender in frontier AI, particularly in open‑weight models. China Daily frames the launch as evidence of “the resources being targeted at the area by developers in China,” noting that Beijing has made AI leadership a national priority.
Moonshot, backed by Alibaba and venture firm HongShan (formerly Sequoia China), has been steadily climbing through Kimi K2, K2.5 and now K3, with each release emphasizing agentic behavior, tool use and multimodal reasoning. By promising open weights at 2.8T scale, the company is offering researchers and domestic firms something neither OpenAI nor Anthropic currently provide at their highest tiers.
For global developers and businesses, K3’s combination of scale, long context, native vision, and eventual open weights makes it a serious alternative to US‑based proprietary models, particularly for use cases that demand local hosting, customization, or deep integration into existing systems.
Moonshot’s new Kimi K3 model therefore matters on several levels: as an engineering experiment in ultra‑large, long‑context design; as a commercial gambit in the high‑end model market; and as a geopolitical marker of China’s ambition to shape the frontier of open AI.
