The benchmark consists of 3,500 VQA pairs across 9 categories, with careful attention to linguistic and cultural diversity: https://www.kimi.com/blog/worldvqa.html (3/3)

Current evaluations often conflate visual knowledge retrieval with reasoning. In contrast, WorldVQA decouples these capabilities to strictly measure "what the model memorizes." (2/3)

Here's our work toward scalable, real-world agentic intelligence. More details in the report 👉https://github.com/MoonshotAI/Kimi-K2.5/blob/master/tech_report.pdf (4/4)

encoder with 4× temporal compression, enabling 4× longer videos in the same context
- Toggle: token-efficient RL, 25–30% fewer tokens with no accuracy drop (3/4)

Quick hits:
- Joint text–vision training: pretrained with 15T vision-text tokens, zero-vision SFT (text-only) to activate visual reasoning
- Agent Swarm + PARL: dynamically orchestrated parallel sub-agents, up to 4.5× lower latency, 78.4% on BrowseComp
- MoonViT-3D: a unified image–video (2/4)

Quick queries cost tiny amounts of tokens, saving your budget for the heavy lifting.
In Short: Your quota goes further. Every bit of compute is now spent exactly on your actual needs.

Try it out here https://www.kimi.com/code (3/3)

No More Waste: In the old system, a simple "Hello World" cost the same quota as refactoring hundreds of lines of code. That’s history.
Precision Billing: With Token-based billing, usage is calculated by actual length. (2/3)

Making your quota go further (1/3)