Limitless: An AI Podcast Kimi K2.5: The Best New Model is Open-Source (Again!)
26 snips
Jan 29, 2026 A new open-source multimodal model can turn screen recordings into full websites in minutes. They cover how the model runs up to 100 sub-agents in parallel and uses an orchestrator with thousands of tools. Conversation highlights cost cuts to pennies per million tokens, video-to-code and creative 3D blueprint uses, and why open weights are shifting developer adoption.
AI Snips
Chapters
Transcript
Episode notes
From Screen Recording To Live Website
- Josh fed Kimi K2.5 a screen recording of Anthropic's website and got a full replica published in about 25 minutes.
- The demo produced a working website without additional prompts, showcasing video-to-code capability.
Multimodal Training Unlocks Video Understanding
- Kimi K2.5 was trained on about 15 trillion tokens spanning text, audio, and visual data.
- Multimodal training lets it understand videos and convert visual input directly into functional outputs like code.
Parallel Sub-Agents Speed Up Workflows
- Kimi K2.5 can spawn up to 100 sub-agents that run tasks in parallel.
- Parallel sub-agents cut execution time significantly, e.g., a 4.5x speedup on complex tasks.