Kimi K2.5 - The biggest open release yet!

Quick Overview

Kimi released K2.5, their first multimodal model, claiming it is the most powerful open-source model to date, outperforming GPT-4o, Gemini 3, and Opus on several benchmarks while offering similar performance to frontier models at a lower cost, and simultaneously introduced the Agent Swarm system powered by K2.5, which orchestrates up to 100 parallel sub-agents.

Key Points: Kimi K2.5 is the first open-weight model to cross a 50 score on the full humanities LSAT, surpassing Gemini 3 Pro's 45.8 score on the same benchmark. The model demonstrates superior performance on agentic use cases like BrowserComp, surpassing even proprietary state-of-the-art models. K2.5 possesses multimodal capabilities including image and video understanding, performing very closely to Gemini 3 Pro in these areas. The release includes Agent Swarm, a system trained with parallel agent reinforcement learning, allowing it to spin up to 100 sub-agents executing parallel work across up to 1,500 coordinated steps. Parallel sub-agents significantly reduce execution time as task complexity increases, showing a minimal increase in time compared to a single agent. K2.5 is a 1 trillion parameter Mixture of Experts (MoE) model with 384 experts, utilizing only 32 billion parameters at a time, and features a context length of 256,000 tokens. In testing, K2.5 successfully generated complex 3JS code for an animation and demonstrated tool usage capability by correcting initial output errors after receiving image feedback.

Context: The Kimi team announced the release of K2.5, marking their entry into multimodal models and positioning it as a leading open-source contender, alongside the introduction of their multi-agent system upgrade called Agent Swarm. This release comes during a busy period for Chinese AI companies, with competitors like Quann releasing Quann 3 Max Thinking and Deepseek dropping Deepseek OCR2, all seemingly aiming to launch products before the Lunar New Year.

Detailed Analysis

Kimi K2.5 is presented as a highly competitive multimodal open-source model, achieving benchmark scores that challenge proprietary leaders like GPT-4o and Gemini 3, particularly excelling in humanities exams and agentic tasks. A key focus for K2.5 is front-end development design, where initial outputs were deemed impressive, though ultimately ranked below Gemini 3 flash in overall design capability. Architecturally, K2.5 is a 1 trillion parameter MoE model with 384 experts, using 32 billion parameters concurrently and supporting a massive 256,000 token context length. The most significant aspect of the release is Agent Swarm, which leverages parallel agent reinforcement learning to orchestrate up to 100 sub-agents, drastically improving performance and reducing execution time on complex tasks compared to single-agent execution. Initial demonstrations showed K2.5 successfully generating and iteratively refining complex 3JS code based on visual feedback, proving its multimodal reasoning and tool utilization capacity, although the full Agent Swarm system is noted as being available only for paid customers.