Kimi K3 Is Live — Moonshot's Flagship Multimodal Reasoner

July 17, 2026

Kimi K3 is in the model picker: Moonshot's 2.8T-parameter open-weight flagship, built for complex coding, knowledge work, and long-horizon agentic runs. It thinks before every answer, reads text and images, and holds a full 1M-token context — the strongest open-weight model Alfrada OS offers today. It bills at the Premium band (6×).

What you can do

• Pick Kimi K3 from the selector — it sits at #2 in the recommended picks, right after GLM 5.2. Routed via OpenRouter (Global).
• Keep whole projects in one pass with a 1M-token context window (1,048,576 tokens) and up to 50K output tokens per response.
• Get always-on deep reasoning — K3 runs with thinking mode permanently enabled at maximum effort; you don't toggle anything.
• Send images, not just text — K3 is multimodal, so screenshots, diagrams, and photos can go straight into the same reasoning pass.
• Use it everywhere agents run — full tool calling, Swarm eligibility, and Auto routing.
• Stay open-weight at the frontier — 2.8T parameters, the largest open-weight model in the catalog, for teams that prefer open models without giving up frontier quality.

Where this shows up

• You have a gnarly multi-file refactor plus the design doc plus the logs. You used to split that across sessions or prune context. Now the whole thing fits in one 1M-token pass with reasoning on.
• You've been using Kimi K2.7 Code as your Moonshot pick. K3 is the newer flagship lane — same family, bigger model, image input, and four times the context.
• You want an open-weight model for a long agentic run. K3 is swarm-eligible and Auto-routable, so it can lead a run or be pinned explicitly.

Try it

• Pick Kimi K3 in the picker, then: "Read this whole repository plus the attached architecture diagram, and propose the refactor plan with the riskiest changes first."
• With Kimi K3 selected: "Work through this contract, the spreadsheet, and these screenshots together, and flag every inconsistency between them."
• With Kimi K3 selected: "Run this multi-step research task end to end and keep the full source material in context instead of summarizing early."

Heads up

• K3 bills at the Premium band (6×) against your plan quota — same token rate as GLM 5.2, GPT-5.6 Terra, and Kimi K2.7 Code. It is not a 12× model.
• Thinking is always on. There is no fast/no-reasoning variant of K3; for quick lightweight turns a Workhorse-band model is the cheaper choice.
• The recommended row grew to four picks — GLM 5.2, Kimi K3, GPT-5.6 Terra, and Claude Opus 4.8. Kimi K2.7 Code leaves the recommended row but stays fully selectable in the picker.