Moonshot AI releases multimodal Kimi K2.5 with Agent Swarm

Background

Kimi K2.5 is the open-weight model that Moonshot AI released at the start of 2026. It is the release that brought together native multimodal input and the Agent Swarm feature ^[1][4]. In practical terms, a user can give the model a mix of text and images, and for a large task the model can hand pieces of the work to sub-agents rather than working through them in one sequence.

What happened

On 27 January 2026 Moonshot AI released Kimi K2.5 as an open-weight model ^[1][4]. The company presented the model as natively multimodal, with a roughly 400-million-parameter MoonViT vision encoder ^[3][2]. It was launched together with Kimi Code, a command-line coding agent ^[2][4].

Key details

The model has 1 trillion total parameters, with 32 billion active for each token. It has 61 layers and 384 experts, with 8 selected per token plus 1 shared expert. It uses MLA attention and supports a 256K-token context window ^[5].

The model's main new feature is Agent Swarm. It breaks a complex task into parallel sub-tasks and assigns them to dynamically created, domain-specific agents, with up to 100 sub-agents working at the same time ^[5][2]. The idea is that a large job can be spread across several agents rather than handled one step after another.

Moonshot's reported scores in Thinking mode include 30.1 on Humanity's Last Exam (HLE-Full) without tools and 50.2 with tools. BrowseComp reached 60.6, and 78.4 when Agent Swarm was used. The model scored 76.8 on SWE-Bench Verified and 50.7 on SWE-Bench Pro ^[5]. SWE-Bench tests whether a model can fix real problems in software projects.

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

Impact

Kimi K2.5 was an open-weight release that combined a large language model with image input and an agent system built for parallel work. For readers, the key idea is that one model can take in several kinds of input and divide a large job among cooperating agents.