moonshotai/kimi-k2-thinking

Kimi K2 Thinking is an open-source, general-purpose agent-level thinking product launched by Moonshot AI, with a core focus on expanding “thinking” as a first-class resource. It’s a deep reasoning model that tackles complex problems through ultra-long thinking tokens and continuous tool calls.

• Industry-Leading Performance: Outperforms GPT-5 and Claude Sonnet 4.5 in benchmark tests such as HLE (Humanity's Last Exam) and BrowseComp, setting new SOTA records.
• Ultra-Fast Experience: Leveraging native INT4 quantization technology, it boosts generation speed by approximately 2x, with API output speeds reaching 60–100 tokens per second—no more long waits for responses.

───────────────────────────────────────────────────────────────────

Core Capabilities

🤖 Model as Agent

No human intervention required—can execute 200–300 consecutive tool calls. While thinking, it uses tools continuously, automatically breaking down complex problems and achieving true autonomous search and browsing.

🧠 Deep Reasoning Engine

Supports Test-Time Scaling, allowing flexible expansion of the thinking token budget. With long-sequence reasoning and reflection mechanisms, it maintains high coherence and accuracy in complex tasks such as mathematics and coding.

🌐 All-Round Information Processing

Supports context windows up to 256k tokens and is compatible with INT4 precision. Optimized for domestic chips, it effortlessly handles long-document reading, complex information gathering, and multilingual code generation.