Kimi K2 put Moonshot AI on the map in July 2025: a 1-trillion-parameter open model that rivalled paid ones at coding and tool use. It grew into a family, from K2 Thinking to K2.8 Preview. Here's what each version did, what still works today and where K2 is heading.
About 32 billion active for each token.
Trained with the MuonClip optimiser, with no loss spikes.
At launch, ahead of GPT-4.1's 54.6% in Moonshot's tests.
Open weights you can download and run.
In the Kimi app, K2.6 is free for everyone in chat and doesn't use credits.
K2.6 and K2.7 Code were closed to new API users on 31 October 2026 and retire by the end of December 2026.
K2.8 Preview runs Kimi Code's default kimi-for-coding ID, and K2.7 Code Highspeed is still offered.
K2, K2 Thinking, K2.6 and K2.7 Code weights stay on Hugging Face for self-hosting.
| Version | Released | What was new | Status, Nov 2026 |
|---|---|---|---|
| Kimi K2 | 11 Jul 2025 | 1T MoE, 32B active, 128K context; Base and Instruct | Retired from the API |
| K2-Instruct-0905 | Sep 2025 | Better coding and front-end work; 256K context | Retired from the API |
| K2 Thinking | 6 Nov 2025 | Reasoning agent, 200–300 tool calls, native INT4 | Retired from the API |
| K2.5 | 27 Jan 2026 | Native vision, Agent Swarm, Kimi Code CLI | Retired 31 Aug 2026 |
| K2.6 | 20 Apr 2026 | Open-weight, stronger long coding and agents; 256K | Free in Kimi chat; API retiring by end Dec 2026 |
| K2.7 Code | Jun 2026 | Coding specialist, about 30% fewer thinking tokens | API retiring by end Dec 2026; Highspeed still in Kimi Code |
| K2.8 Preview | 11 Sep 2026 | Coding model close to K3, 1M context | Current in Kimi Code (kimi-for-coding) |
Details: K2.7 Code, K2.8 Preview and all Kimi models.
| Spec | Kimi K2 (July 2025) |
|---|---|
| Architecture | Mixture-of-Experts with Multi-head Latent Attention (MLA) |
| Parameters | About 1 trillion total, 32 billion active per token |
| Experts | 384 experts, 8 chosen per token plus 1 shared |
| Layers / heads | 61 layers, 64 attention heads |
| Vocabulary | 160K tokens |
| Context | 128K tokens (256K from K2-0905 onwards) |
| Training data | 15.5 trillion tokens of web text, code, maths and knowledge |
| Optimiser | MuonClip (Muon with QK-Clip) |
| Licence | Modified MIT |
| Launch API price | $0.60 input, $0.15 cached, $2.50 output per 1M tokens |
From the Kimi K2 model card and technical report (arXiv 2507.20534).
MuonClip kept K2's training smooth from start to finish.
| Test | What it measures | K2-Instruct |
|---|---|---|
| SWE-bench Verified | Fixing real GitHub issues (agentic, single attempt) | 65.8% |
| SWE-bench Multilingual | Same, across languages | 47.3% |
| LiveCodeBench v6 | Competitive coding | 53.7% |
| Tau2-Bench | Tool use in customer-service tasks | 66.1 |
| ACEBench (En) | Tool and API calling | 76.5 |
| AIME 2025 | Competition maths | 49.5 |
| GPQA Diamond | Graduate-level science | 75.1 |
Moonshot-reported, non-thinking mode. On 17 July 2025, K2 ranked as the top open model on the LMSYS Arena leaderboard. Our independent view: Kimi AI review.
K2 Thinking turned K2 into a reasoning agent that thinks step by step and calls tools for hundreds of steps in a row. Moonshot reported state-of-the-art results at launch.
| Measure | K2 Thinking |
|---|---|
| Humanity's Last Exam (with tools) | 44.9% |
| BrowseComp | 60.2% |
| SWE-bench Verified | 71.3% |
| Sequential tool calls | 200–300 without help |
| Context window | 256K tokens |
| Precision | Native INT4, about 2× faster generation |
Moonshot-reported, all at INT4 precision. Independent leaderboards later showed lower Humanity's Last Exam scores under different setups.
Pick K2.6 in the Kimi app or on kimi.com. It's free in chat and doesn't use credits. How to use Kimi.
kimi-k2.6 costs $0.95 input, $0.16 cached and $4 output per 1M tokens until it retires. Plan your move to K3.1 Pro.
Download the weights from Hugging Face and serve them with vLLM or SGLang on a multi-GPU server.
Keys and billing: Kimi API guide.
| Kimi K2 family | Kimi K3 / K3.1 Pro | |
|---|---|---|
| Parameters | 1T total, 32B active | 2.8T total, 104B active |
| Context | 128K–256K | 1M |
| Vision | From K2.5 | Built in (images and video) |
| API price per 1M (in / out) | $0.95 / $4 (K2.6) | $2 / $10 (K3.1 Pro) |
| Status | Retiring from the API by end of Dec 2026 | Current |
More on the successors: Kimi K3 and Kimi K3.1 Pro.