Kimi K2: Open Agentic Intelligence
by Moonshot AI (Kimi Team)
How a trillion-parameter open-weight MoE was trained end-to-end for agentic tool use.
Overview
Kimi K2 is a Mixture-of-Experts language model with 1 trillion total parameters and 32 billion activated parameters, released with open weights by Moonshot AI. The report details the MuonClip optimizer, which improves on the token-efficient Muon algorithm with a QK-Clip technique for training stability, and reports pre-training on 15.5 trillion tokens with zero loss spike. It then describes a multi-stage post-training pipeline built around a large-scale agentic data-synthesis process that generates tool-use trajectories in simulated and real environments, plus reinforcement learning with verifiable and self-critique rewards. Reported benchmarks include 66.1 on Tau2-Bench, 76.5 on ACEBench (English), 65.8 on SWE-Bench Verified, and 53.7 on LiveCodeBench v6, positioning it among the strongest open-source models for software-engineering and agentic tasks. First submitted July 28, 2025, with a revised version in February 2026.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~1-2 hour read (technical report)
- Provider
- Moonshot AI (Kimi Team)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How a trillion-parameter Mixture-of-Experts model is structured (32B activated of 1T total)
- ✓The MuonClip optimizer and the QK-Clip trick used to keep large-scale pre-training stable
- ✓How agentic tool-use data is synthesized at scale from simulated and real environments
- ✓A multi-stage post-training recipe combining RL with verifiable and self-critique rewards
- ✓How Kimi K2 performs on agentic and coding benchmarks (Tau2-Bench, SWE-Bench Verified, ACEBench)
Highlights
- •First-party account of a frontier-class open-weight model, not a third-party summary
- •Introduces the MuonClip optimizer with a QK-Clip stability technique
- •Deep, concrete treatment of large-scale agentic-data synthesis for tool use
- •Open weights (base and post-trained checkpoints) released alongside the report
Who It’s For
Best For
- ✓ML engineers and researchers training or fine-tuning large MoE models
- ✓Engineers building agentic systems who want to understand tool-use training
- ✓Practitioners evaluating open-weight models for coding and agent workloads
Prerequisites
- •Solid understanding of transformer architectures and LLM pre-training
- •Familiarity with Mixture-of-Experts, optimizers, and RL post-training
- •Comfort reading ML technical reports and benchmark tables
FAQ
What is Kimi K2: Open Agentic Intelligence?
The official technical report for Kimi K2, Moonshot AI's open-weight 1-trillion-parameter Mixture-of-Experts model built for agentic tasks. Written for ML researchers and engineers who want a concrete, reproducible account of how a frontier-class open model is pre-trained and post-trained for tool use.
Is Kimi K2: Open Agentic Intelligence free?
Kimi K2: Open Agentic Intelligence is free to access.
What level is Kimi K2: Open Agentic Intelligence for?
Kimi K2: Open Agentic Intelligence is aimed at a advanced audience. Recommended background: Solid understanding of transformer architectures and LLM pre-training, Familiarity with Mixture-of-Experts, optimizers, and RL post-training, Comfort reading ML technical reports and benchmark tables.
How long does Kimi K2: Open Agentic Intelligence take?
Expect roughly ~1-2 hour read (technical report). Most learners work through it at their own pace.
What will I learn from Kimi K2: Open Agentic Intelligence?
You'll learn: How a trillion-parameter Mixture-of-Experts model is structured (32B activated of 1T total); The MuonClip optimizer and the QK-Clip trick used to keep large-scale pre-training stable; How agentic tool-use data is synthesized at scale from simulated and real environments; A multi-stage post-training recipe combining RL with verifiable and self-critique rewards; How Kimi K2 performs on agentic and coding benchmarks (Tau2-Bench, SWE-Bench Verified, ACEBench).