Beijing-based artificial intelligence start-up Moonshot AI has unveiled Kimi K2, its latest open-source AI model, as part of a strategic push to strengthen its position in the increasingly competitive AI development landscape. The launch is aimed at reinforcing Moonshot’s edge over domestic competitors such as DeepSeek and aligning with the growing industry shift toward open-source AI innovation.
Kimi K2, built on a mixture-of-experts (MoE) architecture, incorporates a staggering 1 trillion total parameters, with 32 billion activated parameters deployed dynamically based on the task at hand. The MoE framework allows for individual subnetworks – or “experts” – to focus on specific data subsets, a technique that reduces pre-training computation costs while accelerating inference performance.
Moonshot has released two distinct open-source versions of Kimi K2. The foundation model, Kimi-K2-Base, is tailored for researchers and developers requiring full control for fine-tuning and bespoke AI applications. Kimi-K2-Instruct, by contrast, has been post-trained for seamless integration into general-purpose chat and autonomous agentic AI systems. Both versions are now available via Moonshot’s web and mobile platforms.
The model’s release underscores a broader move within the Chinese AI sector to embrace open-source development. This approach has been rapidly adopted by peers including Zhipu AI, MiniMax, and Stepfun, as well as tech giants such as Alibaba Cloud and Baidu. Open-sourcing AI models not only enhances development efficiency but also broadens access and encourages adoption within both commercial and academic circles.
Alibaba Cloud’s Qwen family of AI models has emerged as a global leader in the open-source space. According to a February report from Hugging Face, Qwen powers more of the top 10 open-source large language models (LLMs) worldwide than any other group, surpassing Meta Platforms’ Llama in terms of ecosystem scale.
Moonshot’s Kimi K2 arrives in the wake of DeepSeek’s widely noted release of its open-source V3 and R1 models, which demonstrated high performance at significantly lower costs and hardware requirements compared to traditional LLM projects. With Alibaba among its major backers, Moonshot is now positioned to attract increased global attention.
According to Moonshot, Kimi K2 demonstrates “advanced agentic intelligence”, enabling it to execute complex, tool-based workflows autonomously. Examples include producing detailed, interactive salary analyses complete with statistical visualisations and web interface generation. The model can also perform dynamic planning tasks, such as organising a travel itinerary to a London Coldplay concert, by interacting across multiple platforms including Gmail, Airbnb, search engines, and online booking systems.
Moonshot is also planning to introduce advanced model context protocol capabilities, an open standard that allows AI systems to access and utilise external tools and services more effectively.
The company’s API – which is compatible with OpenAI and Anthropic interfaces – is currently priced at 4 yuan (approximately USD 0.56) per million input tokens and 16 yuan per million output tokens.
Shortly after Kimi K2’s launch, OpenAI CEO Sam Altman announced a delay in the release of OpenAI’s own forthcoming open-source model, originally expected the following week, citing the need for additional safety testing.
-SCMP
