Kimi K2.5

What it is

Kimi K2.5 is an open-weight model released by Moonshot AI on 27 January 2026 ^[1][4]. It is natively multimodal and has a roughly 400-million-parameter MoonViT vision encoder ^[5][2]. The model has 1 trillion total parameters, 32 billion active per token, 61 layers and 384 experts, with 8 selected per token plus 1 shared. It uses MLA attention and a 256K-token context window ^[5].

Origins

Kimi K2.5 was built by continual pre-training of Kimi-K2-Base on about 15 trillion mixed visual and text tokens ^[5]. The model therefore starts from Kimi-K2-Base and extends it with visual training data.

It belongs to a series of successive releases, and the model was followed by Kimi K2.6 in April 2026 ^[3][1].

Main features

Kimi K2.5 offers two modes, Thinking and Instant. The Thinking mode works through a problem before answering ^[5].

The model's most distinctive feature is Agent Swarm. It splits a complex task into parallel sub-tasks and runs them with dynamically created, domain-specific agents, up to 100 sub-agents ^[5][2]. The idea is that one large job can be divided among several agents working at the same time, rather than being handled in one long sequence.

The model was launched alongside Kimi Code, a command-line coding agent ^[2][4].

Moonshot's reported Thinking-mode scores include 30.1 on HLE-Full without tools and 50.2 with tools. On BrowseComp it scored 60.6, rising to 78.4 with Agent Swarm. It scored 76.8 on SWE-Bench Verified and 50.7 on SWE-Bench Pro ^[5]. Its multimodal and reasoning results were 78.5 on MMMU-Pro, 86.6 on VideoMMMU, 96.1 on AIME 2025 and 87.6 on GPQA-Diamond ^[5].

The model is released under a modified MIT license, and its technical report is available on arXiv ^[5][6].

Its place in development

Kimi K2.5 pairs a large mixture-of-experts language model with native visual input and agentic features such as Agent Swarm and Kimi Code. The model was followed by Kimi K2.6 in April 2026 ^[3].