Kimi K3.1 Teased Officially; Qwen 4 Linked to Anonymous Arena Model Kalmia

Kimi K3.1 Official Teaser

Moonshot AI released a video showing the Kimi Code interface displaying the model identifier "Kimi K3.1" with a status bar reading "K3.1 thinking: high" . The video demonstrates the model building a 3D scene using Three.js, aligning with earlier community discoveries of a make3d mode in Kimi's frontend code.

Prior signals appeared in late September when the Moonshot API backend showed a kimi-k3-1 identifier. The Kimi official open platform later surfaced a preview labeled "K3.1" or "K3 11111" noting support for 1M token context . On October 21, LuminaBench revealed that Kimi Code's model-switching configuration already contained rules for switching to K3.1, confirming the model's imminent integration.

Anonymous Arena Models: Kibitz and Kalmia

Before the official video, LMSYS Chatbot Arena's Code Arena listed two anonymous models: Kibitz and Kalmia . Community tracker AI Tracker Bot analyzed tokenizer behavior to link them to known vendors.

Kibitz Linked to Moonshot via Tokenizer Tests

The primary evidence for Kibitz being a Moonshot model comes from special token tests . AI Tracker Bot reported that Kibitz failed to correctly reproduce 11 characteristic Moonshot tokens across two test rounds, while seven other tested models handled these strings normally. This anomaly mirrors the behavior of previous Moonshot models.

Historically, before the releases of Kimi K2.5, K2.6, and K3, anonymous models named Kiwido , Kiwire , and Kivine appeared on Arena, reinforcing the naming pattern (K-i-*) and the tokenizer fingerprint as reliable indicators.

Community Benchmarks and Cost Comparison

Users tested Kibitz with creative coding tasks: generating a "pelican on a bicycle" SVG, a pixel-art courtyard, a five-story pagoda, and a Minecraft-style voxel world in a single HTML file. One user reported that Kibitz produced 13,707 blocks in a single Minecraft replication test at a cost of approximately $0.38 , while a comparison model generated only 5,738 blocks for about $5.31 , suggesting a significant efficiency advantage.

Kalmia's Qwen-Like Behavior and 3D Demos

Kalmia exhibits different token anomalies: it reacts abnormally to certain special tokens that also trouble Qwen models. Testers noted Kalmia cannot correctly identify the string .DataGridViewColumnHeadersHeightSizeMode.

Kalmia's public Arena outputs focus on 3D interactive content , including a drivable 3D car and a helicopter with takeoff/landing capabilities. The helicopter demo features exploded views, X-ray mode, and wireframe mode. The car author noted the generated scene still has some issues.

Qwen 4 Technical Background

Alibaba confirmed in September that Qwen 4 is in training . The previously open-sourced Qwen3.8-Flash-Next is described as an early preview of the Qwen 4 architecture. Its technical details reveal the research direction:

Hybrid attention: Gated DeltaNet combined with Qwen Sparse Attention

Gated residuals and N-gram Embedding

Model backbone: 125B parameters , with 6B activated parameters per token

Additional 51B N-gram Embedding parameters

These specifications illustrate Qwen's architectural exploration but do not conclusively prove Kalmia is Qwen 4.

Historical Pattern and Upcoming Release Window

Before Qwen3.8-Max launched, an anonymous model codenamed Kaleb was linked to Qwen by the community. This precedent increases attention on Kibitz and Kalmia, though final confirmation awaits official announcements. With Programmer's Day on October 24 approaching, the community expects answers soon.