Code model with no known source has users pointing to Kimi

A model with no known source has appeared on Arena, the testing site where users compare AI models in head-to-head matchups. It is called Kibitz and has been seen mainly in the coding arena, where models are asked to write code. Many testers guess it is actually Kimi K3.1 from the Chinese company Moonshot AI.

That is only a guess, though. According to the news site TestingCatalog, neither Kimi nor Arena has confirmed that Kibitz and K3.1 are the same model.

Why does the guess point to Kimi?

According to TestingCatalog, Kimi has released a teaser about the upcoming K3.1. Arena's account on X has posted a similar message saying a new model from Kimi is on the way, and asked people to watch how it ranks.

The user Chetaslua was among the first to point to Kibitz. He said the model's JavaScript drawing was very good and fast, and that the result came on the first attempt. Chetaslua also said friends inside Kimi had told him the model will be very efficient. That is secondhand information and has not been confirmed.

What do users' tests show?

The user Lumina shared a 3D helicopter with takeoff and landing, a view where it assembles and explodes, and X-ray and wireframe modes. According to Lumina, it took under 15 minutes to make.

Another user compared Kibitz with Claude Fable 5.1 in a 3D test. That user found Kibitz noticeably faster and judged the result close to Opus 5.5 Medium. The user suspected the model runs at its highest reasoning level by default.

The user Chris shared a comparison in which models had to build themselves in Minecraft. According to the post, GPT-6 Astra used 5,738 blocks at a cost of $5.31, while Kimi K3.1 used 13,707 blocks at a cost of $0.38. Kimi built itself as the blue Kimi Agent mascot, while GPT-6 built a generic robot. The figures come from the post and have not been verified. Chris also claimed that K3.1 will be released before next Tuesday as a cheap, top-tier open source model. Moonshot has not confirmed that.

Take it with a grain of salt

TestingCatalog stresses that these are single attempts in Arena matchups, and that the results should be treated with caution until a model card, an official description of the model, is available.

The predecessor followed the same pattern

According to TestingCatalog, Kimi K3 was seen on Arena in July under the codename Kivine and was released a few days later. It came out on July 16, 2026, with 2.8 trillion parameters and a context window of 1 million tokens. API pricing was $3 per million input tokens and $15 per million output tokens. Open weights, so others can download and run the model themselves, followed on July 27, 2026, under Moonshot's own license.