Kimi K3 Released: 3T Params, 1M Context, Agentic Benchmarks, Pricing & Demos
Referral Link:
Sign up on Kimi and we each get up to 1-Year Membership Credits: https://kimi-bot.com/activities/viral-referral/share?scenario=invite&from=share_poster&invitation_code=N6T4LE
The video reviews Moonshot’s new Kimi K3 model, described as an almost three-trillion-parameter, natively multimodal system with a one-million-token context window, focusing on its release announcement, benchmarks, pricing, and ways to get started. It highlights strong performance across flagship and third-party benchmarks (including DeepSuite, Terminal Bench, FrontierSuite, ProgramBench, and Arena AI), positioning K3 near or above several recent frontier models in specific areas, especially coding and agentic tasks. The script emphasizes K3’s long-horizon agentic focus, architectural notes like Kimi Delta attention residues and a mixture-of-experts setup (activating out of 896 experts), and an example of self-evolving kernel optimization reducing runtime from 283ms to 114ms after 15 hours. Availability is via kimi.com, Kimi Work/Code, the Kimi API, and some third-party providers, with open weights expected July 27; pricing is stated as $3/M input, $15/M output, and $0.30/M cached, alongside demos like recreating macOS in-browser and a rough CS:GO clone.
00:00 Kimi K3 Overview
00:16 Benchmark Highlights
01:06 Agentic Focus
01:23 Architecture And Availability
02:11 Self Evolving Optimization
03:01 Index Speed And Cost
04:39 Token Efficiency
05:40 macOS Web Demo
07:24 Community Reactions
08:10 How To Access K3
08:22 Web App And Plans
09:12 API Credits And Wrap Up
source
