Kimi K2

2025-07-11 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI releases Kimi K2, a MoE model with 1 trillion parameters (32 billion active) with open weights under a modified MIT license, trained with the MuonClip optimizer and designed for agentic uses.

### Context In July 2025, the largest open models (DeepSeek-V3, Qwen3) remain well below closed models in agentic coding. Moonshot AI extends Moonlight (Muon) to the trillion-parameter scale. An update, Kimi-K2-Instruct-0905, later raises the context to 256K and improves agentic coding: it is grouped in this entry. ### What the model brings - Open MoE with 1 trillion parameters and 32 billion active, pre-trained on 15.5T tokens. - MuonClip optimizer: Muon combined with qk-clip, which rescales the query and key matrices to control the explosion of attention logits; the vendor claims training without instability spikes. - Optimization for tool use: the model is described as a “reflex” model without long thinking, which acts rather than answers. - Two releases: Kimi-K2-Base (foundation) and Kimi-K2-Instruct (post-trained). - Technical report on arXiv 2507.20534 (July 28, 2025). ### Architecture and specifications MoE with 1T parameters in total and 32B active, 61 layers (including 1 dense), 384 experts of which 8 selected per token and 1 shared, MLA attention, SwiGLU activation, 160K vocabulary, 128K context (256K for the 0905 update). Text only. License: modified MIT (free commercial use; obligation to display “Kimi K2” in the interface beyond 100 million monthly active users or $20 million in monthly revenue). Weights: huggingface.co/moonshotai/Kimi-K2-Instruct. ### Cost and access | Item | Price (USD per million tokens) | |---|---| | Input (cache miss) | 0.60 | | Input (cache hit) | 0.15 | | Output | 2.50 | Pricing recorded from secondary sources (the official pricing page has removed the kimi-k2 series, discontinued on May 25, 2026 according to the platform documentation). Self-hosting hardware: not disclosed in the model card. ### Benchmarks Kimi-K2-Instruct (without extended thinking) compared with closed models without extended thinking. Scores published by the vendor in the model card. | Benchmark | Kimi K2 | Claude Sonnet 4 | Claude Opus 4 | GPT-4.1 | |---|---|---|---|---| | SWE-bench Verified (agentic, 1 attempt) | 65.8 | 72.7 | 72.5 | 54.6 | | LiveCodeBench v6 | 53.7 | 48.5 | 47.4 | 44.7 | | AIME 2025 (Avg@64) | 49.5 | 33.1 | 33.9 | 37.0 | | GPQA-Diamond | 75.1 | 70.0 | 74.9 | 66.3 | | MMLU | 89.5 | 91.5 | 92.9 | 90.4 | | Tau2 telecom | 65.8 | 45.2 | 57.0 | 38.6 | Source: Hugging Face card for Kimi-K2-Instruct (vendor measurements; some competitor values taken from their publications). ### Versus closed models On this table, K2 beats GPT-4.1 on five of the six tests listed (all except MMLU) and surpasses Claude Sonnet 4 and Opus 4 on LiveCodeBench, AIME 2025 and Tau2 telecom. It remains behind Claude Sonnet 4 and Opus 4 on agentic SWE-bench Verified (65.8 versus 72.7 and 72.5) and on MMLU. The comparison covers closed models without extended thinking, not reasoning models. On price, the API is at 0.60 USD input and 2.50 USD output (secondary sources), and the weights allow self-hosting, which meets a sovereignty need, at the cost of a trillion-parameter infrastructure. ### Key takeaways Kimi K2 brings open models into the trillion-parameter category with a permissive commercial license. For enterprises, it offers a credible open foundation in agentic coding, provided they have suitable infrastructure.

Tags : LLM, Open Source, Moonshot AI, Kimi, MoE, MIT modifiée, Agentique