Kimi K3.1 from Moonshot AI and Qwen4 from Alibaba have entered anonymous testing on Arena, the platform where users compare models blind, and an official launch could come this week.
Max For AI reported that the two models are in their final stages before release. Two anonymous models with codenames, Kibitz and Kalmia, have appeared on the platform. The account @aitrackerbot analysed their tokenizers and concluded that Kibitz shows characteristics consistent with Kimi, while Kalmia appears to come from Qwen. Both teams ran intensive benchmark tests in recent days, and the presence of the models on Arena suggests an imminent release.
For Qwen4, Alibaba officially announced the family at the Yunqi Conference in September, with Max, Flash and Plus variants and a 27-billion-parameter model for local use. For Kimi, K3 is already out as an open-weight model of 2.8 trillion parameters, with native vision and a 1-million-token context window.
A post by Lumina on X compared the two directly: same prompt, a 3D helicopter with takeoff and landing, 3D assembly and exploded views, with both taking roughly 15-20 minutes. Lumina says the Kimi K3.1 version was made in under 15 minutes and includes X-ray and wireframe modes. The post drew about 76,800 views.
The open question is how Arena results translate into real availability. Kimi K3 has so far been held back by compute limits, slow speeds and low subscription quotas, so the K3.1 release will be watched closely on price and access.
No comments yet. Be the first!