Where it stands
Kimi has not announced or documented K3.1. Earlier reports rest on a LuminaBench registry probe and on @AiBattle_'s claim of two SGLang commits, one tied to a Deepinfra deployment. A new X post from @HarshithLucky3 says K3.1 is being tested on Arena as "kibitz". None of these has been confirmed by Kimi, and official Kimi documentation or an announcement would be needed to confirm K3.1.
An X poster links a model called "kibitz" on Arena to Kimi K3.1
@HarshithLucky3 says on X that Kimi K3.1 is in testing on Arena (arena.ai) under the name "kibitz". He says he sees it in almost every comparison when he tries Carbon outputs on Arena.
He also describes the model's design as the same as Astra's, with what he calls Sonnet-like taste in 3D work. He says that if the model is distilled from Astra and Sonnet, it would give Astra-level output at 10x less cost.
Our read: the post is a user's observation in a public comparison interface, which is a different kind of evidence from the earlier registry and commit reports. The post text does not explain how the "kibitz" label was tied to Kimi K3.1, and the distillation point is his speculation. Kimi has not announced K3.1 or confirmed this model, and no official Kimi page has been found that names it.
Post on X reports K3.1 in two SGLang commits and a Deepinfra deployment
@AiBattle_ says on X that Kimi K3.1 was recently spotted in two SGLang commits, and that one of them references a K3.1 deployment already running on Deepinfra. The post text names no commit hashes or Deepinfra listing, so the commits and the deployment are unverified.
The post also speculates that K3.1 could appear on the Arena this week, following the naming pattern of earlier Kimi models, and cites K2.5 and K2.6 as examples.
Our check: Kimi's official Kimi-K3 repository on GitHub uses kimi-k3 as the API model name and does not mention K3.1. That fits with Kimi not having announced K3.1, but it does not confirm or rule out the commits or deployment the post describes.
Kimi K3.1 slug found in a backend registry
A report from LuminaBench says the kimi-k3-1 model slug survived a synchronization and backend probe, suggesting that an entry with that name was callable through a router.
In a post on X, @LuminaBench says a synchronization process pulled models from an official product list, tested the resulting entries and removed models rejected by the backend. The kimi-k3-1 entry remained callable after that process.
Image credit: @LuminaBench on X
LuminaBench clarified that the entry did not come from Kimi’s code; it was observed through a router performing the synchronization. LuminaBench suggested that the model could be in testing but said it did not know when it might be released.
0 comments
No approved comments yet. You can start the conversation.
Leave a comment