Descriptions:
Matthew Berman covers the release of Kimi K2.5, a natively multimodal open-weights model from Chinese AI lab Moonshot AI, positioned as a state-of-the-art competitor for agentic tasks and front-end coding at significantly lower cost than closed frontier models.
Kimi K2.5 was pre-trained on approximately 15 trillion mixed visual and text tokens and introduces native self-directed agent swarm capability: the model can decompose complex tasks and coordinate up to 100 sub-agents executing up to 1,500 parallel tool calls. Kimi reports a 4.5x speedup and 80% reduction in end-to-end runtime versus single-agent operation on complex tasks. Benchmark highlights include BrowseComp at 74.9 (beating GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro), SWE-Verified at 76.8 (trailing GPT-5.2 and Claude Opus 4.5 but beating Gemini 3 Pro), and MMMU-Pro vision at 78.5 (beating Claude Opus 4.5). The model is trained using a novel parallel agent reinforcement learning (PARL) method.
Berman highlights cost-performance ratio as the model’s defining edge—on several benchmarks it matches or exceeds GPT-5.2 at a fraction of the price. Demos include website recreation from screenshots and multi-step visual path-finding using BFS code execution. The model is available at kimi.com and as downloadable open weights for local deployment.
📺 Source: Matthew Berman · Published January 27, 2026
🏷️ Format: Review
