Code Arena users report Kaleb shows Qwen-like token quirks

• Kaleb appeared as an anonymous Arena test rather than a public Qwen release: Lentils80 reported the stealth name on LMArena, while Arena's leaderboard policy says providers can test unreleased models behind unique anonymous labels.
• The Qwen attribution rests on token behavior: AiBattle_ said Kaleb and several Qwen models mishandle prompts such as PostalCodesNL in similar ways.
• The best visible capability signal was a browser 3D task: synthwavedd reported strong 3D performance, and Lentils80's attached run showed a 167,736-voxel scene at 60 fps.
• A second anonymous model also appeared: synthwavedd said torenia-alpha entered Code Arena with Kaleb, with too little testing to identify its origin.

Arena's leaderboard policy says unreleased models can be tested behind unique anonymous labels until enough votes stabilize. The Code Arena help page says the coding surface supports fullstack apps with databases, APIs, and backend deployments. Alibaba's Model Studio release notes list June Qwen3.7 Max and Plus snapshots, which leaves Kaleb outside the public Qwen naming shown there.

Kaleb repeatedly identifies itself as Claude during an LMArena test.

Kaleb showed up under a stealth name in Lentils80's report, which described it as a possible Qwen 3.8 or Qwen 4 model. The same report said users leaned toward Qwen based on Qwen-specific output signals and political-question probes, while still calling the identity unconfirmed.

The self-ID was the noisy part. In the captured run, Kaleb repeatedly says it is Claude.

Token quirks are the strongest Qwen clue in the evidence pool. According to AiBattle_'s comparison, the fingerprint was:

• PostalCodesNL can be treated as unrelated words.
• The same token can make the prompt behave as empty.
• The screenshot compared Kaleb with odd outputs from Qwen3.6 Plus, Qwen3.7 Plus, and Qwen3.7 Max.

It remains a behavioral fingerprint rather than a confirmation.

synthwavedd reported strong 3D-task performance from Kaleb and called it likely Qwen's next flagship. The visible Kaleb artifact is a browser-rendered voxel scene labeled 167,736 voxels, 60 fps, and 65 draws.

The official Code Arena help page frames this as the expected surface: users can build apps and websites inside an AI development environment with databases, APIs, and backend deployments.

The thread had a material correction. synthwavedd's first post called Kaleb likely Kimi's next flagship; the corrected post changed the attribution to Qwen.

In a reply, synthwavedd said they were tired and had fixed it.

torenia-alpha arrived alongside Kaleb, according to synthwavedd's Code Arena report. Its origin was left open because synthwavedd said they had not tested it enough.

The cutoff clue was separate, with the same report putting both models' knowledge cutoffs in late 2025 or early 2026.