Alibaba-backed Moonshot AI launches a new open-source model, rivaling DeepSeek
What's the story
Chinese artificial intelligence (AI) start-up, Moonshot AI, has launched a new open-source AI model, named Kimi K2. The move comes as part of a larger trend among Chinese companies to open-source their AI models. This is in stark contrast to many US tech giants like OpenAI and Google, who keep their most advanced models proprietary. The new model boasts improved coding capabilities and excels at general agent tasks and tool integration.
Model features
Kimi K2 challenges DeepSeek's V3
The Kimi K2 model is designed to break down complex tasks more effectively. Moonshot AI claims that this new model outperforms mainstream open-source models in some areas, including DeepSeek's V3, as well as the capabilities of leading US models like those from Anthropic in certain functions, such as coding. This puts Moonshot AI back in the race against other Chinese companies that have released similar open-source models. The company hopes to reclaim market presence with the new launch.
Performance
New model features 1 trillion parameters
Kimi K2 has a total of one trillion parameters, with 32 billion activated at a time using a mixture-of-experts architecture. Two versions are being released: a foundation model for researchers and developers, and an instruction-tuned version tailored for chat and autonomous agent use cases. The model's key strength lies in its optimization for "agentic" abilities, enabling it to autonomously use tools, generate and run code, and carry out complex multi-step tasks without human input.
The scores
Take a look at the benchmark scores
In benchmark evaluations, Kimi K2 scored 65.8% on SWE-bench Verified, a demanding software engineering benchmark, surpassing most open-source models and matching the performance of some proprietary ones. On LiveCodeBench, widely regarded as the most realistic coding benchmark, it achieved 53.7% accuracy, significantly outperforming DeepSeek's V3 (46.9%) and OpenAI's GPT-4.1 (44.7%). Most notably, it reached 97.4% on MATH-500, outpacing GPT-4.1's 92.4%, indicating that Moonshot may have unlocked a core breakthrough in mathematical reasoning that has eluded larger and better-funded rivals.
Twitter Post
Official announcement from the company
🚀 Hello, Kimi K2! Open-Source Agentic Model!
— Kimi.ai (@Kimi_Moonshot) July 11, 2025
🔹 1T total / 32B active MoE model
🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models
🔹Strong in coding and agentic tasks
🐤 Multimodal & thought-mode not supported for now
With Kimi K2, advanced agentic intelligence… pic.twitter.com/PlRQNrg9JL
Scenario
Delivering results at a fraction of the cost
Moonshot AI is delivering results with a model that costs only a fraction of what incumbents spend on training and inference. While OpenAI pours hundreds of millions into compute, Moonshot has found a far more efficient route to the same goal. The company has open-sourced Kimi K2 and is simultaneously offering an aggressively priced API. At $0.15 per million input tokens (cache hits) and $2.50 per million output tokens, Moonshot undercuts OpenAI and Anthropic while matching, or even beating, their performance.
Company history
What is Moonshot AI?
Founded in 2023 by Tsinghua University graduate Yang Zhilin, Moonshot AI is one of China's leading AI start-ups. The company has the backing of internet giants like Alibaba. It gained popularity in 2024 for its long-text analysis capabilities and AI search functions. However, its position has slipped after DeepSeek launched low-cost models like the R1 model that disrupted the global AI industry. Moonshot's Kimi app, once third in monthly active users last August, slipped to seventh by June, per aicpb.com.