On July 11, 2025, a Beijing lab pushed a model with one trillion parameters onto a public repository. Within days, that model — Kimi K2 from Moonshot AI (月之暗面) — was sitting at the top of the open-source leaderboard on LMArena, ahead of every other openly released model. The move reopened a question the industry keeps asking: how far can an open-weight model get when it is built specifically to act, not just to talk?
Kimi K2 is not a reasoning model in the step-by-step "think hard" sense. It is an agentic model (智能体模型) — tuned so that, given a set of tools, it figures out how to chain them together and finish the job.
Key takeaways
- Kimi K2 is a Mixture-of-Experts (MoE) model with 1 trillion total parameters and 32 billion activated per token, released as open weights (开源权重) under a Modified MIT license.
- It is positioned as an "open agentic (智能体) intelligence" model, optimized for tool use and multi-step coding rather than long chain-of-thought reasoning.
- At launch it reached number one among open models on LMArena and scored state-of-the-art results for open weights on SWE-bench Verified, Tau2, and AceBench.
- Moonshot open-sourced two checkpoints — Kimi-K2-Base and Kimi-K2-Instruct — and exposed an API at roughly US$0.55 (HK$4.3) per million input tokens and US$2.20 (HK$17) per million output tokens.
What the model actually is
Kimi K2 uses a sparsely activated MoE design. Out of 1 trillion total parameters, only 32 billion fire for any single token. That keeps the cost of running it far below a dense model of comparable scale while preserving a very large capacity. The architecture is built on 384 routed experts with 8 selected per token plus one shared expert, Multi-head Latent Attention (MLA), and 61 layers. Context length is 128K tokens.
The model was pre-trained on 15.5 trillion tokens using a custom optimizer Moonshot calls MuonClip. The lab says the optimizer was the real story: as human-quality training data becomes a bottleneck, squeezing more intelligence out of each token matters as much as raw compute. MuonClip adds a qk-clip technique that rescales query and key weight matrices to stop attention logits from exploding during trillion-parameter training — a stability problem that had bitten earlier large runs.
The agent angle, not the reasoning angle
Most 2025 flagship releases leaned into "reasoning" — models that deliberate before answering. Kimi K2 takes a different bet. It is a reflex-grade model: it answers without a long internal monologue, but it is tuned to call tools and execute. In Moonshot's own demos, Kimi K2 books a Coldplay tour through 17 tool calls spanning search, calendar, Gmail, flights, Airbnb, and restaurants; it refactors a Flask project into Rust while running benchmarks; it automates Minecraft mod development in JavaScript by managing rendering, running tests, and debugging until green.
The training pipeline backs this up with two pieces of engineering: large-scale synthetic agentic tool-use data, and a general reinforcement learning phase with a self-evaluation signal. The result is a model that treats "use the tools I gave you" as a first-class skill rather than an afterthought.
Benchmarks and reception
On the LMArena public arena, organized by Stanford and LMSYS, Kimi K2 became the highest-ranked open-source model and the fifth-ranked model overall, behind only four closed systems. Nature's website noted the release on July 16, 2025, describing it as "another DeepSeek moment" for Chinese AI.
On coding and agentic benchmarks, the numbers Moonshot reported include:
- SWE-bench Verified: 65.8% (single attempt), the best among open-weight models at launch.
- Tau2 Retail: 70.6%.
- AceBench: 76.5%.
- MATH-500: 97.4%; MMLU: 89.5%; GPQA-Diamond: 75.1%.
- AIME 2024: 69.6% (avg@64).
By late July, the weights had passed 195,000 downloads on Hugging Face, and on OpenRouter the model topped the trending chart with more than 57.5 billion paid tokens called in a week. Those are usage signals, not benchmark scores, but they show developers actually pulling the model into production.
How to get it
Two checkpoints are public: Kimi-K2-Base, the raw pre-trained foundation for fine-tuning, and Kimi-K2-Instruct, the post-trained version for chat and agentic use. The API mirrors OpenAI and Anthropic formats, so existing tooling can switch with minimal changes. The weights remain freely downloadable even after Moonshot retired the original first-party 0711 checkpoint from its own API in May 2026.
Honest limitations
- Kimi K2 is text-only. It has no vision or audio; Moonshot said thinking and visual understanding would come later.
- Because it is a non-reasoning model, it is weaker than dedicated reasoning models on competition math that needs long deliberation (its AIME 2025 sits at 49.5%, below reasoning-focused peers).
- Benchmark figures are largely self-reported by Moonshot under its own prompts and attempt counts; treat them as a strong starting point, not a settled ranking.
- Running the full 1T model still needs a serious GPU cluster; the open weights help, but this is not a laptop model.
What readers can do now
- Try it through the API or a third-party host (OpenRouter) on one real tool-calling task — booking, data cleanup, or code refactor — and compare latency and cost against a closed model.
- Pull Kimi-K2-Instruct from Hugging Face and run a quantized version locally if you have a single capable GPU, to see agentic behavior without sending data to a server.
- If you build agents, prototype with its tool-use format and measure whether "act without overthinking" fits your workflow better than a reasoning model.