Moonshot AI drops Kimi K3 — its most capable model yet, with 1M context
Moonshot AI quietly released Kimi K3 on Thursday evening — the company’s most capable model to date, and one that brings the Chinese AI lab’s flagship series to an interesting inflection point.
The new model supports up to a 1 million token context window, matching the industry’s top tier for long-context reasoning. It’s being positioned as particularly strong on three fronts: programming, gaming and 3D tasks, and knowledge-intensive queries. That’s a deliberate choice — coding and spatial reasoning are where some of the most competitive AI benchmarks live right now.
Kimi Code, the company’s developer platform, now offers two models (K3 and K2.7 Code) across three model IDs. The tier structure breaks down like this:
- Andante tier: K3 not supported
- Moderato tier: 256K context window
- Allegretto and above: up to 1M context
As of publication, Moonshot AI has not published official blog posts, pricing details, or benchmark results on its website or social channels. Technical specs — parameter count, architecture details, inference cost — remain unannounced.
The K3 launch follows a brisk cadence for Moonshot. The lab released Kimi K2.6 in April 2026, which brought upgrades in code generation, long-horizon task execution, and agent orchestration — and notably, that model was open-sourced. Before that, the original Kimi K2 arrived in July 2025, a 1-trillion-parameter MoE model with 32 billion active parameters, built for agent-heavy workloads. A September 2025 update (K2 0905) stretched context from 128K to 256K and delivered faster API speeds.
The question now is whether K3 will follow the open-source trajectory of K2.6, or stay proprietary. Given the competitive landscape — DeepSeek, Alibaba’s Qwen, ByteDance’s Doubao, and Zhipu AI are all pushing hard on both performance and openness — Moonshot’s pricing and licensing strategy for K3 could be the real story.
For now, the model is live but mostly unspoken for. That silence might not last long.