Kimi K2.5: agentic vision in the most powerful open-source model

Moonshot AI has launched Kimi K2.5, which according to published benchmarks is presented as the most powerful open-source model to date. The model incorporates vision, coding, and multi-agent system processing capabilities, with a context window of 256,000 tokens. It is available through Kimi.com, the Kimi app, Moonshot's API, and Kimi Code.

K2.5's development has been based on continued training with approximately 15 trillion tokens combining visual and textual data. This multimodal architecture enables the generation of complete interfaces from conversations, implementing interactive designs and complex animations. The model can reconstruct complete websites from videos, solve visual puzzles by marking the shortest path through code, and perform autonomous visual debugging, as demonstrated when translating the aesthetics of Matisse's work into a web interface by iterating on its own output.

The system called "agent swarm" (multi-agent) represents a shift in scaling strategy. K2.5 can self-direct up to 100 sub-agents executing parallel workflows with a maximum of 1,500 coordinated tool calls. The technology uses Parallel-Agent Reinforcement Learning (PARL) to decompose tasks into parallelizable subtasks executed concurrently. This approach reduces execution time by up to 4.5 times through a metric called Critical Steps that measures latency following the critical path concept in parallel computing. In internal evaluations, Agent Swarm mode has demonstrated an 80% reduction in execution time for complex tasks.

In office productivity, K2.5 shows improvements of 59.3% in the AI Office benchmark and 24.3% in General Agent compared to K2 Thinking. In software engineering tasks it achieves 76.8% on SWE-Bench Verified. The model coordinates multiple tools to generate documents, spreadsheets, PDFs, and presentations, supporting tasks such as adding annotations in Word, building financial models with pivot tables, and writing LaTeX equations, scaling up to documents of 10,000 words or 100 pages.

The platform offers four operating modes: K2.5 Instant, K2.5 Thinking, K2.5 Agent, and K2.5 Agent Swarm, the latter in beta phase. For software engineering, Moonshot recommends Kimi Code, its open-source tool that runs from the terminal and can integrate with environments like VSCode, Cursor, or Zed.

Key points

• Moonshot AI presents Kimi K2.5 as the most powerful open-source model according to published benchmarks
• Excels in vision-based programming, generating code from images and videos, with 76.8% on SWE-Bench Verified
• Incorporates an "agent swarm" (multi-agent) system that coordinates up to 100 sub-agents in parallel, reducing execution time by up to 4.5 times
• Offers advanced office productivity capabilities, generating documents, spreadsheets, and presentations with improvements exceeding 50% compared to K2
• Available in four different modes, including Agent Swarm in beta, and complemented with Kimi Code for software development