Moonshot Kimi K2.6 & K3: the open-source agentic coder

Among the 2026 Chinese open-source models, Kimi from Moonshot AI is the specialist — less a general chatbot, more an engine for autonomous coding agents that plan, edit files, run tools and keep going for dozens of steps. It earned that reputation with Kimi K2, extended it with K2.6, and pushed it to a new scale with K3.

What Kimi is, in one line

Kimi is an open-weight mixture-of-experts (MoE) language model tuned for tool use and long-horizon agent tasks. MoE means the network has a very large total parameter count but only routes each token through a small subset of “experts,” so it is far cheaper to run than its raw size suggests. (For the underlying idea, see our parameters & scaling laws guide.)

K2.6: closing the frontier gap

Kimi K2.6 is a roughly trillion-parameter MoE that, on 2026 coding arenas, ranked among the top open models and traded blows with Qwen 3.6 Max and DeepSeek V4 — falling only just behind the best closed systems. Its strength is agentic: given a repository and a goal, it can decompose the task, call tools, and iterate. That is a harder skill than one-shot answers, and it is what makes Kimi useful for real engineering work rather than demos.

K3: the largest open-weight model yet

In July 2026 Moonshot released Kimi K3, a reported ~2.8-trillion-parameter model with a 1M-token context window, described at launch as the most capable Kimi and one of the largest openly released models to date. The huge context matters for agents: it can hold an entire codebase, a long task history and tool outputs in working memory at once — the thing that usually breaks long agent runs. (Why context is a token budget: our tokens explainer.)

How the agentic training works

Models like Kimi are not just pre-trained on text; they are post-trained heavily on tool-use and multi-turn trajectories — sequences where the model must call a function, read the result, and decide the next action. This is closer to reinforcement-learning-style training than plain next-token prediction, and it is why an agentic model behaves so differently from a base model on the same benchmark. Our fine-tuning, distillation & RLHF guide covers the family of techniques involved.

Licence and access

Kimi ships as open weights under a permissive (modified MIT-style) licence, so you can download it from Hugging Face and self-host, or call it through Moonshot's Kimi API or aggregators like OpenRouter. Because K3 is enormous, most teams will start on the hosted API and only self-host K2-class models where hardware allows.

Where Kimi fits for a Malaysian team

Kimi shines when the job is build-and-ship software with an agent: internal tools, data pipelines, migrations, test generation. Pair it with an orchestration layer — the skills we teach in Mastering Claude & Multi-Agent Orchestration transfer directly — and with automation platforms like n8n to wire the agent into real systems. For rapid app prototyping with AI, see AI Vibe Coding.

Limitations to keep in mind

Scale is not free: K3's size means real GPU cost to self-host, and a 1M context is expensive to fill on every call. As with any model, verify agentic claims on your own repositories — agent benchmarks and your messy production codebase are different animals. And always pin a specific model version in production, because these families iterate monthly.

Official resources & downloads

Go straight to the source — official sites, model cards and weight downloads:

• Moonshot AI — official site
• Kimi — try it (chat)
• Kimi model weights on Hugging Face