Kimi K3.1 Revealed: Moonshot’s Next Model + Qwen4 Near

Moonshot AI just named its next model — in a 24-second video. On October 23, 2026, the company’s official Kimi account posted a retro-computer-style clip in which a Kimi Code terminal reads “Model: Kimi K3.1”. That single frame is the first official confirmation that Kimi K3.1 exists, and reports say a launch could come this week.

Per ChainCatcher’s report, the video shows the model writing code for a complex 3D scene using Three.js. KuCoin’s flash news and Binance Square, citing ChainCatcher and MetaEra, put the release at roughly 02:01–02:26 UTC on October 23. No specs, no price — just a name and a clear signal of direction: agent-ready code generation with visual output.

The breadcrumbs that led here

Since September the community has been tracking the trail:

• September 19: Kimi’s verified Zhihu account posted 135 digits of pi, no caption. The community decoded it as “3.1” — an unverified but telling interpretation.
• Late September: developers found the identifier “kimi-k3-1” on Moonshot’s API platform, per ChainCatcher.
• October 21: K3.1 appeared in Kimi Code’s model-switching configuration, in a screenshot by LuminaBench — a pre-launch clue, not a product announcement.

Confirmed vs. unconfirmed: what we actually know

Qwen4: what Alibaba actually said

At the Yunqi (Apsara) Conference on September 22, 2026, Alibaba officially announced that Qwen4 is “currently in training,” and projected Qwen 4.5 and Qwen 5 at 5–10 trillion parameters — a roadmap, not a spec sheet.

Four tier names — Max, Flash, Plus, and a 27B model — circulated from a captured stage slide: Max as the capability ceiling, Flash and Plus as efficiency-and-price tiers, 27B for local deployment. Those names don’t appear in Alibaba’s written release, so treat the roster as reported, not documented.

Developers have one concrete foothold: Qwen3.8-Flash-Next, shipped August 26, which Alibaba’s own model card describes as an experimental preview of the Qwen4 architecture. For scale, the previous flagship shipped open weights as Qwen3.8-2.4T-A95B — 2.4 trillion parameters, 95B active. Alibaba builds flagships you can download, and Qwen4 will likely follow that playbook. Details live on Qwen’s official site.

The Arena mystery: “Kibitz” and “Kalmia”

Now the speculative part — clearly labeled. Two anonymous models, codenamed “Kibitz” and “Kalmia,” are being tested blind on LMArena, where users compare models without vendor labels. Community tokenizer analysis suggests Kibitz looks like a Kimi model and Kalmia like a Qwen model. That mapping is unverified speculation, but it’s smart speculation — and both teams have reportedly run intensive benchmark sessions in recent days.

Lumina Bench also posted a same-prompt demo between the two — a 3D helicopter with takeoff, landing, and exploded views — drawing about 76,800 views. Third-party demo, not an official benchmark, and Max For AI reports both models are in their final stages before release — reports, not confirmations.

How to profit from the Kimi K3.1 and Qwen4 wave

Open-weight frontier models are how independent builders make money — you run them, fine-tune them, and sell on top:

1. Build Three.js workflows now. The reveal video is advertising 3D code generation. Product configurators, interactive data visualizations, and game prototypes are sellable this week — start on current Kimi models and swap in K3.1 on launch day.
2. Test the Qwen4 architecture preview. Qwen3.8-Flash-Next is available and previews Qwen4’s architecture. Build against it now for a head start on everyone who waits for the announcement.
3. Line up GPU access. K3 shipped open weights; if K3.1 follows, a ready eval set and fine-tuning pipeline turn launch day into ship day.
4. Sell the migration. Kimi K3’s reported launch API pricing was $3/$15 per million input/output tokens — a fraction of closed frontier models. Walk into client calls with that math and you’re the expert in the room.

Watch: Apsara Conference 2026, where Qwen4 entered training

FAQ

Is Kimi K3.1 officially released?

No — only the name has been officially revealed. Reports say a launch could come this week, but neither Moonshot nor Alibaba has confirmed a date.

What did the official video show?

A 24-second retro-computer-style clip where the Kimi Code terminal reads “Model: Kimi K3.1” while generating complex Three.js 3D scene code.

Do we know Kimi K3.1’s specs or price?

Nothing official. Its predecessor, Kimi K3, was a 2.8-trillion-parameter open-weight model with a 1-million-token context window and native vision.

Is Kimi K3.1 open weight?

Unconfirmed. The expectation comes from K3’s open-weight release — Moonshot hasn’t published K3.1 licensing.

What has Alibaba confirmed about Qwen4?

Only that it is “currently in training” (September 22, 2026) and that Qwen 4.5/5 are projected at 5–10 trillion parameters. Tier names (Max/Flash/Plus/27B) come from a reported stage slide.

Are “Kibitz” and “Kalmia” really Kimi K3.1 and Qwen4?

Unverified community speculation from tokenizer analysis. The anonymous Arena models exist; the mapping is an educated guess.

When will the models launch?

Reports say this week is possible. Treat that as rumor until Moonshot AI or Alibaba announces.

Sources: ChainCatcher, KuCoin, Binance Square, alextech.ai, kimi.ai, qwen.ai, lmarena.ai. All figures attributed; speculation labeled as such.

Keep Reading

• NVIDIA Nemotron 3.5 Transcribe Turns Audio Into Answers
• OpenAI Plans to Buy Google AI Chips, Report Says
• Microsoft MAI Image 2.7 Beats Rivals on Hard Edits