Chinese AI startup Moonshot AI officially released the third-generation flagship model Kimi K3 today (17th), with a total parameter of 2.8 trillion (Trillion) and a maximum of 1 million Token (1,048,576) ultra-long context window. It is known as the open source AI model with the largest number of parameters in the world. The model has been launched simultaneously on Kimi Code, Kimi app (including iOS), Kimi Work and Kimi API, and has promised to open the weight before July 27 so that developers around the world can download and deploy it themselves.
Kimi K3 is Moonshot AI’s third-generation new architecture following the K2 series (K2, K2.5, K2.6, K2.7 Code). It adopts Mixture-of-Experts (MoE) design and has a total parameter count of 2.8 trillion, but the official number of active parameters for a single inference has not been announced. Moonshot AI is headquartered in Beijing and is famous for its Kimi chat robot and K2 series models. After K2 was open sourced under a Modified MIT license last year, it once became one of the most powerful open source weight models in 2025. The company completed a US$500 million Series C round of financing in January this year, with a valuation of US$4.3 billion. The funds were clearly used for K3 development and computing expansion. CEO Yang Zhilin also publicly stated at the beginning of the year that K3 is the next major version of the company.
Delta Attention: A new hybrid attention mechanism
The biggest technical highlight of Kimi K3 is Moonshot’s self-developed Delta Attention architecture, a hybrid linear attention mechanism that can achieve up to 6.3 times decoding acceleration in a context scenario of millions of Tokens. The traditional Transformer attention mechanism will incur a computational cost of O(n²) in long context scenarios, causing memory and time overhead to increase sharply with the input length. Delta Attention significantly reduces the computational burden through linearized attention calculations, which is particularly important for Agent scenarios that need to process complete code libraries or extremely long files. It also makes the context of millions of Tokens more usable and economical in practice. According to official documents, K3 supports automatic contextual caching, and the input price of a cache hit is only one-tenth of the normal input.
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional… pic.twitter.com/eFHEbdxn3P— Kimi.ai (@Kimi_Moonshot) July 16, 2026
In addition, Moonshot also incorporates the new technology Attention Residuals, which improves training efficiency by approximately 25% at an additional cost of less than 2%. This technology is very practical in large-scale MoE models. Training costs have always limited the scale development of open source models. It can improve training efficiency with almost no increase in budget, which means that the same computing power can train stronger models. K3 also supports reasoning_effort parameter setting (max level is currently available), allowing developers to dynamically adjust the reasoning depth of the model to strike a balance between simple tasks and complex reasoning.
Two variants and pricing strategies
Kimi K3 is launched in two variants: K3 Max focuses on chat and agent tasks and is very suitable for daily conversations and program assistance; K3 Swarm Max is designed for large-scale parallel processing scenarios and is suitable for enterprise-level application scenarios that need to activate a large number of Agents at the same time. Both versions support native multimodality (image and video understanding), automatic contextual caching (no manual configuration required), structured JSON output, and tool integration capabilities, including custom tools and dynamic tool loading. K3 also adds a new tool_choice constraint function, which allows developers to precisely control when a model uses a specific tool, which is very practical in complex Agent workflows.
In terms of pricing, K3 charges $3 per million input tokens, $15 for output, and only $0.30 for cache hits, which is exactly the same as the officially announced pricing. The context window is exactly 1,048,576 tokens (approximately 1 million), which can accommodate complete project code and a large amount of contextual information in coding and knowledge work scenarios. This price positioning is similar to Anthropic’s Sonnet series and slightly higher than the input rate of GPT-5.6 Terra ($2.50). However, some developers pointed out on Hacker News that inference efficiency may be more important than single Token pricing. If Kimi K3 needs to spend more inference Tokens to achieve the same results, the actual cost of use may not be lower than that of competitors. The company also simultaneously launched a monthly subscription plan of 199 yuan, which is an entry-level option. API stored value can enjoy 10% to 30% rebate before August 11.
The strength of the world’s largest open source model
Kimi K3 is the open source AI model with the largest number of parameters in the world. Moonshot AI stated that “more parameters represent a higher upper limit of capabilities and can provide smarter performance” and claimed that the overall intelligence level of K3 is close to the world’s leading closed-source model. Estimates from independent review platform Artificial Analysis show K3 Max performing around the Opus 4.8 / GPT-5.5 level, but still lags behind Anthropic’s latest Fable 5 in some arena cues.
From the perspective of iteration speed, Moonshot AI’s research and development pace is quite amazing. The K2 series debuted in July 2025, less than a year after K2.7 Code, and K3 officially debuted only three months after K2.6 was open source.
The key significance of open source ecology
The timing of Kimi K3’s open source is also worthy of attention. Just the day before, xAI was forced to open source Grok Build due to privacy disputes; OpenAI launched the Codex Micro exclusive hardware keyboard on the same day. Within a week of the explosion of AI coding tools, Moonshot chose to grab the attention of the developer community with the model with the largest number of open source parameters. After the weight is opened on July 27, third-party developers will be able to deploy, fine-tune, and audit K3 on their own. This is far more attractive to enterprise users who value data privacy and customization needs than the closed API solution. According to the information that OpenRouter has put on the shelves, K3 can be accessed through the OpenAI compatible API format, which lowers the conversion threshold for developers.
Conclusion
With 2.8 trillion parameters, million Token context, and Delta Attention architecture, Kimi K3 directly challenges the flagship status of OpenAI GPT-5.6 and Anthropic Fable 5. Although it is still not as good as the top closed-source models in some evaluation indicators, considering its open source positioning and aggressive API pricing strategy, K3 is a very competitive choice for developers who need self-deployment or large-scale calls. After the weight is opened on July 27, the next thing worth observing is its actual performance in community deployment and fine-tuning scenarios.
Source: KOCPC Chinese