Kimi K2

Kimi K2 is a Mixture-of-Experts (MoE) model from Moonshot AI, released in July 2025 with an accompanying arXiv technical report. It has 1.04 trillion total parameters with 32 billion active per token, and Moonshot positions it as "open agentic intelligence" — an open-weight model designed for agentic and coding workloads without relying on extended chain-of-thought reasoning at inference time. Its principal documented contributions are the MuonClip optimizer with QK-Clip, an industrial-scale agentic data synthesis pipeline, and joint reinforcement learning with self-critique rubric rewards. (Source: Kimi K2 Technical Report)

Capabilities and benchmarks

Kimi K2 is documented as an open-weight MoE built for agentic and coding workloads. It shares an architectural family with DeepSeek-V3 — both use Multi-head Latent Attention and a sparse-MoE design — but diverges on several design axes: 50% more total experts (384 vs. 256), half the attention heads (64 vs. 128), and a different optimizer (MuonClip). (Source: Kimi K2 Technical Report)

Benchmark scores as reported in the technical report:

Training and architecture

The model's documented design choices center on training stability, sparsity, and long-context serving efficiency. (Source: Kimi K2 Technical Report)

The MuonClip optimizer combines the token-efficient Muon optimizer with QK-Clip, a per-head query/key weight-rescaling rule, γ_h = min(1, τ/S_max^h), applied after each optimizer step. Without it, attention logits exceed magnitude 1000; with τ = 100, they stabilize and decay. K2 trained on 15.5T tokens with "no observable spikes." A sparsity-scaling analysis found that increasing sparsity lowers both training and validation loss: sparsity 48 (384 experts / 8 active) reduces FLOPs by 1.69× at loss 1.5 compared with sparsity 8. The attention-head count reflects a stated trade-off — halving heads to 64 costs 0.5–1.2% validation loss but cuts 128k-context inference FLOPs by 83%, an optimization for long-context agentic serving.

The training data and post-training process are built around agentic use. An agentic data synthesis pipeline draws on 3,000+ real MCP tools plus 20,000+ synthesized domain tools to generate thousands of synthetic agents with rubric-based tasks, producing stateful multi-turn trajectories filtered by an LLM-as-judge, complemented by real code and SWE execution sandboxes. Post-training uses joint reinforcement learning with self-critique rubric rewards: a Verifiable Rewards Gym spanning math/STEM, instruction-following, faithfulness, coding, and safety, with the critic continuously refined by on-policy RLVR rollouts. The technical report is among the more detailed public accounts of synthetic agent-trajectory generation at industrial scale (Agentic AI).

Licensing and openness

Accounts of the K2 weight license differ across sources. The arXiv technical report describes the weights as released under CC BY-NC-ND 4.0 — non-commercial, no derivatives (Source: Kimi K2 Technical Report). Contemporaneous coverage of the Hugging Face weight release, however, describes a Modified MIT License — permissive and commercial-use-friendly, with an added attribution condition requiring prominent "Kimi K2" display for large-scale commercial deployments (Source: interconnects.ai; venturebeat.com). Under the CC BY-NC-ND reading, K2 would be more restrictive than the Apache-2.0-style licenses used by DeepSeek — publicly inspectable but not freely redistributable or commercially usable; under the Modified MIT reading it sits close to the permissive end of the spectrum. Either way, the case illustrates how "open" is being operationalized heterogeneously by Chinese labs, relevant to Open-Source AI / Open-Weight Models debates about what "open" should mean for weights.

Reception and context

Alongside DeepSeek-V3 and DeepSeek-R1, K2 is cited as further evidence that Chinese labs are among the most public contributors to MoE training-stability and efficiency literature (US-China AI competition). Its shared MLA-MoE design with DeepSeek-V3 suggests a converging Chinese architecture family distinct from frontier US dense-decoder and sparse-expert practices.

Related models

Kimi K2 Thinking, a reasoning variant with the same 1T-total / 32B-active MoE architecture, was released November 6, 2025 under the Modified MIT License. It interleaves extended thinking with tool calls across long sequential chains, and contemporaneous coverage described it as the leading open-weight model on several agentic and reasoning benchmarks at release (Source: simonwillison.net; venturebeat.com).

A successor, Kimi K2.6, was released in April 2026, scoring 58.6 on SWE-bench Pro at roughly 5–6× lower cost than Claude Opus 4.7, with a demonstrated 12-hour autonomous agent run. (Source: artificialintelligencemade-simple.com)

The K2 generation was succeeded on July 16, 2026 by Kimi K3, a ~2.8-trillion-parameter model on a new attention architecture (Kimi Delta Attention and Attention Residuals) with a 1M-token context window, which Moonshot reports achieves roughly 2.5× better overall scaling efficiency than K2 (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)).

Relationships

• instance-of: General-Purpose AI (GPAI)
• related: DeepSeek-R1, Qwen3, Agentic AI, Kimi K3