Before the end of this month, the two most anticipated flagships in the domestic large model camp — Kimi K3.1 and Zhipu GLM-5.5 — are gradually surfacing, and Kimi K3.1 has quietly started running tests on the Arena battleground under the codename Kibitz.
Several netizens, following clues such as the tokenizer, have basically confirmed that Kibitz is Kimi K3.1. Earlier leaks sketched a clear direction for the upgrade: focusing on long-horizon autonomous tasks, strengthening gaming and 3D code capabilities, and filling in computer operation skills — an overall approach clearly modeled on the "top-tier playbook" of frontier models like OpenAI.
Today, quite a few netizens actually got their hands on it and gave it a spin, and the feedback was explosive — whether it was tricky images like pelicans or 3D animation generation, K3.1 showed strength capable of going toe-to-toe with GPT-6 Astra. According to Arena convention, large models tested on the battleground are usually officially released a week later, so there is still hope of an official announcement and launch of Kimi K3.1 by the end of the month.
Market expectations for it are not low: K3.1 will likely become a new benchmark for domestic AI large models and has the capability to squeeze into the global TOP3; on the AA leaderboard, it is hoped to become the first domestic model to break 50 points, ideally reaching around 55 or even 60 points. However, beyond performance, K3.1 still has to make up ground in usability and cost-effectiveness — the previous generation K3, strained by tight compute resources, ended up rather slow with limited plan quotas, and over the past two months its buzz was overshadowed by fast-and-cheap models like DeepSeek V4.1 Flash. As the domestic flagship prepares to go head-to-head with Astra, posting pretty scores alone is not enough — whether developers can use it smoothly and affordably is the real key to turning the tables.