Kimi K2.5

2026-01-27 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI has released Kimi K2.5, a natively multimodal open model of 1 trillion parameters, extending K2 on 15 T image and text tokens, with an agent swarm of up to 100 sub-agents.

### Context In January 2026, the leading closed models (GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro) lead in vision and coding. Kimi K2.5 takes the K2 and K2 Thinking base and makes it multimodal, while pushing multi-agent orchestration further. ### What the model brings - Continued pretraining on about 15 T mixed image and text tokens on top of Kimi-K2-Base, with no trade-off between vision and text according to the vendor. - Coding with vision: reconstructing a website from a video and autonomous visual debugging. - Agent Swarm (research preview): up to 100 sub-agents and 1,500 coordinated tool calls, trained with Parallel-Agent Reinforcement Learning (PARL); execution time reduced by up to 4.5 times compared with a single agent, according to the vendor. - Instant and thinking modes, native INT4 quantization carried over from K2 Thinking. - Technical report on arXiv 2602.02276. ### Architecture and specifications 1T-parameter MoE, 32B active, 384 experts, 256K context, MoonViT vision encoder, text, image and video as input. License: modified MIT (display "Kimi K2.5" beyond 100 million monthly active users or $20 million of monthly revenue). Weights: huggingface.co/moonshotai/Kimi-K2.5. Available on Kimi.ai, the app, the API and Kimi Code. API identifier: kimi-k2.5 (discontinued on August 31, 2026 according to the platform). ### Cost and access | Item | Pricing (USD per million tokens) | |---|---| | Input | 0.60 | | Input (cache hit) | 0.10 | | Output | 3.00 | Pricing from secondary sources (the official pricing page no longer lists K2.5). Hardware: not disclosed. ### Benchmarks Scores published by the vendor (Hugging Face card). The source's asterisks indicate re-runs by the vendor for competitors. | Benchmark | K2.5 (Thinking) | GPT-5.2 (xhigh) | Claude 4.5 Opus | Gemini 3 Pro | |---|---|---|---|---| | HLE-Full, with tools | 50.2 | 45.5 | 43.2 | 45.8 | | SWE-bench Verified | 76.8 | 80.0 | 80.9 | 76.2 | | GPQA-Diamond | 87.6 | 92.4 | 87.0 | 91.9 | | BrowseComp | 60.6 | 65.8 | 37.0 | 37.8 | | Terminal Bench 2.0 | 50.8 | 54.0 | 59.3 | 54.2 | | MMMU-Pro | 78.5 | 79.5 | 74.0 | 81.0 | | AIME 2025 | 96.1 | 100 | 92.8 | 95.0 | Source: Hugging Face card for Kimi-K2.5 (vendor measurements). ### Versus closed models With tools, K2.5 leads on HLE (50.2 versus 45.5 for GPT-5.2 and 43.2 for Claude 4.5 Opus) according to the vendor. It remains behind in coding (76.8 on SWE-bench Verified versus 80.0 and 80.9) and on GPQA-Diamond against GPT-5.2 and Gemini 3 Pro. In vision, it is close to GPT-5.2 on MMMU-Pro (78.5 versus 79.5) and ahead of Claude 4.5 Opus. The vendor emphasizes the cost-performance ratio ("at a fraction of the cost") without quantifying the gap in the blog post. The choice between open and closed comes down to self-hosting and the permissive license more than to a clear performance advantage. ### Key takeaways Kimi K2.5 makes the K2 family multimodal and introduces the agent swarm as an execution mode. It remains below the best closed models in coding, but is the closest to their level in agents and vision among the open models of the time, according to the vendor.

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