Summary
Moonshot AI’s Kimi K2.5, released in late January 2026, is the subject of this detailed walkthrough from David Ondrej. The Chinese AI lab — founded by Yang Zhilin (formerly of Google Brain, Carnegie Mellon PhD, co-author of key transformer papers) and backed by Alibaba and Tencent with over $2 billion raised at a $4 billion valuation — positions K2.5 as a direct competitor to frontier closed-source models at a fraction of the cost: $0.60 per million input tokens and $3 per million output tokens versus Anthropic Claude Opus 4.5’s $5 and $25 respectively.
The headline technical feature is a built-in agent swarm: K2.5 can autonomously spawn and coordinate up to 100 parallel sub-agents without user configuration, a capability baked into the model architecture through parallel agent reinforcement learning rather than implemented as a prompting trick. The model is also natively multimodal, trained on 15 trillion mixed text-and-image tokens, with a mixture-of-experts architecture activating 32 billion parameters. Complex tasks reportedly complete up to four times faster than sequential alternatives.
The video also addresses a controversy around K2.5 occasionally identifying itself as Claude, speculating on whether Anthropic’s outputs were used as synthetic training data or, in a more concerning scenario, whether model weights were leaked. Setup walkthroughs cover kimmy.com, the Kimi Code IDE extension (a Claude Code competitor), and API access via OpenRouter.
📺 Source: David Ondrej · Published January 28, 2026
🏷️ Format: Review
