Kimi, Qwen, Kibitz and Kalmia together drew about 4,200 posts in our X sample, including 915 on Oct 22 alone. Two code names showed up on the Arena leaderboard, and people want to know whose models they are.
Short answer: nobody official has said. This article separates what's reported from what's confirmed, then shows how to read a leaderboard so the next mystery model doesn't fool you.
What's been reported
ChainCatcher, as carried by Gate on Oct 22, citing the account Max For AI, reports:
- Moonshot AI's Kimi K3.1 and Alibaba's Qwen 4 are "in their final stage before release".
- Both are being tested anonymously on Arena, with launch expected the week of Oct 19 (as reported on Oct 22).
- Accounts tracking AI releases matched Kibitz to Kimi and Kalmia to Qwen, based on tokenizer analysis.
AI Tracker, in a write-up on Kibitz, says it detected the model on Oct 21 at 8:44 PM EDT. Its test: Kibitz garbled all 11 test tokens that only Moonshot's tokenizer splits in a particular way. That's strong evidence Kibitz runs on Kimi's tokenizer. AI Tracker's own conclusion is careful: "most likely" the K3.1 that had recently surfaced in sglang code, "but the version isn't confirmed." (AI Tracker's page was unreachable when we last checked on Oct 22, so we're relying on our earlier read.)
What's confirmed
Not much:
- Moonshot hasn't announced K3.1. Its latest official model is Kimi K3: 2.8T parameters, a 1M-token context, and API pricing of $3 input, $15 output and $0.30 cache-hit per million tokens, per its K3 blog.
- Alibaba's last official update, on Sept 22, 2026 at its Apsara Conference, said Qwen 4 was "in training". Its headline also said future Qwen models would scale to 10 trillion parameters.
So Kalmia being Qwen 4 fits the timeline, but nobody official has confirmed it.
How tokenizer fingerprinting works
Every model family splits text into tokens its own way. Some strings get split oddly by one tokenizer and cleanly by others. Ask an anonymous model to repeat those strings, and the way it stumbles reveals the tokenizer underneath.
It's clever, but it has limits:
- It identifies the tokenizer family, not the exact version. That's why AI Tracker won't say "K3.1" for certain.
- Labs can share or license tokenizers, so a match is strong evidence, not proof.
- A model saying "I'm made by X" proves even less. Models often name the wrong maker.
How to read an Arena leaderboard
Arena-style leaderboards rank models by blind head-to-head votes. They're useful, with caveats:
- Check the vote count. A new anonymous model with few votes has a wide confidence range. An early #1 can drop a lot.
- Look at the category. "Overall" mixes chat, coding and creative writing. If you build apps, the coding or WebDev view matters more.
- Votes measure preference, not correctness. Confident, well-formatted answers win votes even when they're subtly wrong.
- The pre-release build may not ship. Labs test variants. The launched model, its price and its limits can differ.
- Wait for the model card. Context length, pricing, license and whether the weights are open decide whether a model is useful to you.
What to do until they launch
Nothing, unless you like testing early. When the official releases land, run your own evals. Our guides on picking a coding agent and what to check before trusting AI-written code explain how. If either model ships with open weights, check our VRAM guide before you try to run it at home. A 2.8T-parameter K3-class model isn't a laptop model.
