moonshotai/kimi-k2-instruct

Kimi-K2-Instruct is a trillion-parameter MoE language model launched by Moonshot AI, primarily positioned as a enterprise-grade “reflective” agent (Agentic) foundation, specifically designed for high-concurrency, low-latency tool calls and complex task execution.

• The world’s first open-source trillion-parameter MoE: With a total of 1 trillion parameters, it dynamically activates only 32 billion parameters during inference, striking a balance between top-tier model performance and exceptionally high inference efficiency.
• Leading coding capabilities: Achieved an astonishing score of 65.8% in the SWE-bench Verified test, surpassing GPT-4.1 and capable of autonomously completing the entire workflow of code generation, debugging, and optimization.
• Enterprise-level cost reduction: Compared to traditional dense models, its computational costs are reduced by more than 35%; in e-commerce customer service scenarios, labor costs can be cut by up to 40%, and response speed is improved by a factor of three.
• Real-world long-text handling: Natively supports a context window of 128K, and combined with the MLA mechanism, it efficiently handles ultra-long document analysis and complex codebase comprehension.
• Out-of-the-box compatibility: Requires no complicated configuration to support multi-turn function calls, perfectly compatible with mainstream inference engines such as vLLM and SGLang, as well as OpenAI/Anthropic API formats.

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

Core Capabilities

🚀 Ultimate efficiency: Adopting an innovative MoE architecture, it boasts 1 trillion parameters yet dynamically activates only 32 billion “expert” parameters per inference, achieving a win-win situation of “large model performance” and “low cost”.

🛠️ All-purpose tool: Specifically designed for Agent scenarios, natively supporting web browsing, code execution, and mathematical software calls; significantly outperforms comparable open-source models in Tau2 tool usage evaluations.

💻 Code master: Sets new open-source records in both LiveCodeBench and SWE-bench—capable not only of writing code but also of autonomously fixing bugs and refactoring projects like an experienced engineer.

🧠 Precise decision-making: Trained with the MuonClip optimizer, its logical reasoning and mathematical computation abilities (AIME 2024 score: 69.6) rank among the industry’s top tier, excelling at handling complex financial and research tasks.

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

Related Evaluations

“Global Debut of Billion-Parameter Model Kimi K2: Code Capability Evaluation—Is It Truly Impressive or Just Hype?”