--- title: "Kimi K3 Is China's 2.8 Trillion Parameter Open AI Model" description: "Moonshot AI's Kimi K3 is a 2.8 trillion parameter open-weight model with a one million token context. Here is what it is, how it performs and how to use it." author: "Chandra Shekhar" published: 2026-10-31 updated: 2026-10-31 canonical: https://answers.mindstick.com/blog/1495/kimi-k3-is-china-s-2-8-trillion-parameter-open-ai-model category: "Kimi" tags: ["Kimi", "Moonshot AI", "open models"] reading_time: 5 minutes --- # Kimi K3 Is China's 2.8 Trillion Parameter Open AI Model On 16 July 2026, Chinese company Moonshot AI released Kimi K3, a model with 2.8 trillion parameters that the company describes as open-weight. It quickly drew attention for strong results in coding and web browsing tasks, and for a price far below many American frontier models. If you have heard the name and want to understand what Kimi K3 actually is, here is a clear overview. ## Who makes Kimi Moonshot AI is a Beijing-based AI company founded in 2023. Its Kimi assistant became popular in China for handling long documents, and the company gained international attention with its earlier open models, including the Kimi K2 series. Kimi K3 is its most ambitious model so far. ## What 2.8 trillion parameters means Parameters are the internal values a model learns during training. More parameters generally allow a model to store more knowledge and handle more complex patterns, though training quality matters just as much as size. Kimi K3 uses a design called mixture of experts. Instead of using all 2.8 trillion parameters for every word it processes, the model contains many specialised sub-networks, called experts, and activates only a small number of them for each token. This gives the model the knowledge capacity of a very large system while keeping running costs closer to those of a much smaller one. According to the TECHSY review of the launch, Kimi K3 also uses a new attention mechanism called Kimi Delta Attention, a hybrid linear attention design that Moonshot says decodes up to 6.3 times faster at very long context lengths. ## Key specifications Based on published reports at launch: - **Context window:** one million tokens, enough for very long documents or large codebases. - **Inputs:** text, images and video. - **Reasoning:** always on, with a maximum reasoning level available at launch. - **API pricing:** about $3 per million fresh input tokens, $0.30 per million cached input tokens and $15 per million output tokens, according to reports. - **Access:** the Kimi web app, Kimi's coding and work tools, and an API that is compatible with OpenAI's software development kit. Epoch AI, an independent research group that tracks AI models, lists Kimi K3 with roughly 2.0 × 10^25 operations of training compute and ranked it eighth of the models it tracked at the time of release. ## How it performs Moonshot reported strong results on several tests. On BrowseComp, a test of how well an AI can find hard-to-locate information by browsing the web, Moonshot reported a score of 91.2. On long-horizon coding tasks, the company's own tables placed it ahead of some leading American models. Epoch AI lists scores of 93 percent on GPQA Diamond, a test of graduate-level science questions, and 60 percent on ARC-AGI-2, a test of abstract reasoning. Two cautions apply. First, many launch benchmarks were run by Moonshot in its own testing environments and had not been independently verified at release. Second, other models still led on some tests. Reports noted that Anthropic's Claude Fable 5 and OpenAI's models led on certain coding and reasoning benchmarks. As with any model, real-world testing on your own tasks matters more than headline scores. ## Open weights, with a caveat Open-weight means the model's trained parameters are released so that anyone can download and run the model on their own hardware. This is different from closed models, which you can use only through the company's app or API. At launch, the weights were not yet public. Moonshot said they would be released on 27 July 2026. Reports also noted that the exact licence terms should be checked on the official model card. Before using Kimi K3 commercially, read the licence carefully. There is also a practical point. A 2.8 trillion parameter model, even with mixture of experts, needs serious hardware to run. Most individuals and small companies will use it through the API or hosted providers rather than running it themselves. ## Why it matters Kimi K3 is part of a broader pattern. Chinese AI labs, including DeepSeek, Alibaba's Qwen team and Moonshot, have been releasing increasingly capable open models at low prices. This puts pressure on American companies on price and gives developers worldwide more choice. For developers and businesses, open models offer several advantages: the option to run them privately, freedom to fine-tune them for specific tasks, and independence from any single provider's pricing or policies. ## Things to consider before using it **Data privacy.** If you use Moonshot's own app or API, your data is processed on its servers. Read the privacy policy, and avoid sending sensitive personal or business data unless you are comfortable with how it is handled. Running the open weights yourself, or through a provider you trust, avoids this concern. **Content rules.** Hosted services from Chinese companies follow Chinese regulations, which can affect how they answer politically sensitive questions. **Organisational policy.** Some companies and government bodies have rules about using AI models developed in particular countries. Check your organisation's policy. **Verification.** Treat launch benchmark claims as a starting point and test the model on your own work. ## How developers in India can try it The easiest route is the API, which works with the OpenAI software development kit, so existing code can often switch to Kimi K3 by changing the base URL, API key and model name. Developers interested in self-hosting should wait for the official weights and licence, and check whether their hardware or cloud budget can support a model this large. ## The bottom line Kimi K3 is a very large open-weight model from Moonshot AI with a one million token context window, strong reported results in coding and web browsing, and competitive pricing. It shows how quickly Chinese labs are closing the gap with American frontier models. If you build with AI, it is worth testing, especially for agentic and coding work, while keeping an eye on licence terms, privacy and independent evaluations. --- Original Source: https://answers.mindstick.com/blog/1495/kimi-k3-is-china-s-2-8-trillion-parameter-open-ai-model Copyright © MindStick Software Pvt. Ltd. This Markdown version is provided for developers, AI systems, and offline reading.