Mystery Models Kibitz and Kalmia Appear on CodeArena, Suspected to Be Kimi K3.1 and Qwen4
机器之心 · wechat · 2026-10-23
Two anonymous codenames — Kibitz and Kalmia — have appeared on Arena.ai's CodeArena blind coding leaderboard, with the community suspecting they are Moonshot's Kimi K3.1 and Alibaba's Qwen4 respectively.
- Benchmarks in the wild: users have coaxed the models into producing SVGs, web pages, and interactive 3D scenes (a medieval pixel castle, a drivable 3D car, helicopters). Some think Kibitz could set a new open-source record, and Kalmia may rival GPT 6.1 SolUltra.
- Cost: community estimates put Kibitz's token cost for the same task at 1/14 of GPT-6 Astra's.
- Kimi evidence: a request to a config endpoint on Kimi's official domain returned a switch target "to":"k3.1" (later withdrawn), and AITrackerBot found Kibitz garbles the same 11 fingerprint tokens that identify Moonshot models.
- Qwen4 direction: Qwen3.8-Flash-Next was officially positioned as an experimental preview of Qwen4's architecture, combining Gated DeltaNet with Qwen Sparse Attention, gated residual connections, and N-gram Embedding for long-context, cross-layer, and parameter-scaling efficiency.
Anonymous Arena appearances typically precede launches by days, so new releases from both labs may be imminent.
More from Models
- Study: 96% of ChatGPT memories are saved without users asking — petesena · 2026-10-24
- OpenAI to showcase the next era of cyber defense on November 12 — imjustnewatai · 2026-10-24
- Unitree's Wang Xingxing: GPT-6 Astra handles robot painting and Go with no specialized training — Afinetheorem · 2026-10-24
- Uncensored Qwen3.8-27B GGUF model trends on Hugging Face — ttindall · 2026-10-24
- Viral claim: GPT can signal across sandboxes via timed file writes; ex-OpenAI safety VP jokes about ASI — TheMoonMidas · 2026-10-24
- Jack Dorsey touts full-duplex realtime voice model trained on a single GB200 NVL72 — TinfoilTricorn · 2026-10-24