Anonymous Models Kibitz and Kalmia Surface in Code Arena: Kimi K3.1 and Qwen 4 Suspected

Two unknown participants, Kibitz and Kalmia, have been spotted over the past two days in Code Arena, the section of Arena.ai (formerly LMArena) where models compete on programming tasks. No lab has yet claimed these models as its own. Still, many people who have analyzed the anonymous models' responses link them to new Chinese models. The names being floated are Kimi K3.1 and Qwen 4.

Two unfamiliar models showing up in the same section at the same time is uncommon. Usually a blind test features a single codenamed candidate, and the community spends weeks guessing who is behind it. Here there are two candidates at once, so theories about their authors spread quickly.

How blind testing works

Arena.ai builds its rankings on pairwise comparisons. A user submits a task, receives two answers from unnamed models, and picks the better one. The names are revealed only after the vote. For coding, there is a dedicated Code Arena mode that evaluates programming ability specifically.

Labs use this mechanism as a proving ground before release. The model gets live feedback from thousands of people, while competitors and the press cannot see the brand and have no chance to form a bias. One well-known example: in 2024, OpenAI tested a model on LMArena under the name im-also-a-good-gpt2-chatbot, and it later shipped as GPT-4o. An anonymous debut in the arena has long been an established practice and often precedes an official announcement.

Why suspicion has fallen on Kimi and Qwen

The community's conclusions rest on indirect signals: response style, reasoning manner, code formatting, and reactions to specific prompts. Such signals can support a hypothesis, but they are not proof. The names Kimi K3.1 (developed by Moonshot AI) and Qwen 4 (Alibaba's model family) remain commenters' guesses for now. The companies themselves have confirmed nothing.

Something else is worth noting: both product lines have long competed in open and commercial coding models, and coding has become the main arena for comparisons. If the anonymous models do turn out to be Chinese, the simultaneous launch of new generations from two prominent families would increase pressure on Western providers of coding assistants.

What is known so far

• Platform: Arena.ai, formerly LMArena
• Section: Code Arena, blind pairwise comparisons on programming tasks
• Models: Kibitz and Kalmia, which appeared in the past two days
• Community hypotheses: Kimi K3.1 and Qwen 4
• Status: no official statements from the developers

How to track anonymous models in the arena

While the labs stay silent, it is more useful to watch practical signals. If you work with coding models, compare Kibitz and Kalmia on your own tasks, provided the arena serves them to you in your battles. Working repositories, long context, and tool use will tell you more than the overall ranking.

Rankings for such models are unstable in the first days: votes are few, confidence intervals are wide, and a model's position can shift noticeably within a week. Teams choosing a model provider for production are better off waiting for the official release with its description, license, and API pricing. Treat anonymous results as an early indicator, not a reason to change your stack.

Official announcements from Moonshot AI and Alibaba will show whether the guesses were right. If they were, the names will most likely be revealed alongside the release and the publication of benchmarks.