AI
Moonshot AI Releases Open-Source Kimi K2 Thinking Model
Moonshot AI releases new open-source AI model, attracting widespread attention
3 MIN READ
By Timmy
Published:
Chinese artificial intelligence startup Moonshot AI has released Kimi K2 Thinking, an open-source AI model that outperforms leading proprietary systems on key reasoning benchmarks, the company announced on November 6, 2025.
The model, available through Hugging Face and Moonshot's API platform, represents a significant advancement in open-weight AI capabilities. Kimi K2 Thinking scored 44.9% on Humanity's Last Exam with tools, 60.2% on BrowseComp, and 71.3% on SWE-Bench Verified—benchmarks where it surpasses OpenAI's GPT-5 and Anthropic's Claude Sonnet 4.5 in several categories.
"Kimi K2 Thinking sets a new state-of-the-art on Humanity's Last Exam, BrowseComp, and other benchmarks by dramatically scaling multi-step reasoning depth," Moonshot AI stated in its technical documentation. The model maintains stable tool use across 200–300 sequential calls, a capability previously seen only in closed systems.
The release uses a Modified MIT License that permits full commercial use, with one requirement: products serving over 100 million monthly users or generating $20 million monthly revenue must display "Kimi K2" in their interface.
Kimi K2 Thinking employs a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion active parameters per token. Its native INT4 quantization reduces the model size to 594GB—down from the original K2's 1.03TB—while doubling inference speed and halving GPU memory usage.
The model features a 256,000-token context window and demonstrates particular strength in agentic coding tasks, achieving 61.1% on SWE-Multilingual and 47.1% on Terminal-Bench. It can autonomously browse websites, execute code, and iterate on solutions across hundreds of steps without human intervention.
Moonshot AI, backed by Alibaba, first released the K2 series in July 2025. The Thinking variant adds step-by-step reasoning capabilities while preserving the distinctive style that made earlier versions popular among developers.
Users can access Kimi K2 Thinking through kimi.com, platform.moonshot.ai, or host it locally using vLLM, SGLang, KTransformers, or TensorRT-LLM inference engines. The model currently lacks vision capabilities, though Moonshot plans to add this feature in future updates.
Industry analysts note the release intensifies competition in the AI sector. "The buzz around this model so far is very positive. Could this be the first open weight model that's competitive with the latest from OpenAI and Anthropic?".
The model's training cost approximately $4.6 million, according to a source cited by CNBC. Moonshot AI has not independently verified this figure.
In internal tests, Moonshot identified limitations including potential token generation issues during hard reasoning tasks and performance degradation with one-shot prompting for complex software projects. The company stated it is working to address these concerns in future releases.
