Kimi K3: the largest open-weight model ever shipped

Kimi K3 is Moonshot AI's flagship, released July 16, 2026, and a milestone for the open-weight world: at 2.8 trillion parameters it is the largest open-weight model ever shipped — roughly triple its predecessor K2.6's 1T. It pairs a 1M-token context window with multimodal reasoning (text + image input) and a training focus Moonshot has doubled down on since K2: large-repo navigation and long-horizon agentic work, where the model must hold a goal across hundreds of files and many tool calls. Pricing lands at $3 per million input tokens and $15 per million output — a big step up from K2.6's $0.60/$2.80 and squarely in premium territory alongside Claude Sonnet 5's standard sticker. K3 is the first open-weight model priced and positioned as a direct flagship peer rather than a budget alternative. The case for Kimi K3: agentic coding across huge codebases, open-weight requirements at frontier quality, and Chinese-language strength. The case against: no native web search, and one-shot reasoning where GPT-5.6 Sol and Claude Fable 5 keep the edge.

Specs

Best at

• Large-repo navigation — the headline strength: finding and editing the right code across hundreds of files in a 1M-token window.
• Long-horizon agentic work — Moonshot's K-series specialty since K2: plan-act-observe loops that don't lose the goal.
• Open-weight at frontier scale — organizations that require inspectable or self-hostable weights get flagship quality for the first time.
• Multimodal agent flows — reads screenshots, diagrams, and UI mockups mid-loop.

Where it loses

• No native web search — pair it with a search-capable council seat (Grok 4.5, Gemini) for anything time-sensitive.
• One-shot reasoning — GPT-5.6 Sol and Claude Fable 5 still produce cleaner single-pass answers on hard prompts.
• Price vs its own sibling — K2.6 at $0.60/$2.80 remains the value pick when you don't need K3's scale.
• Self-hosting practicality — open weights at 2.8T parameters are a statement more than a deployment plan for most teams.

Kimi K3 vs Kimi K2.6

K2.6 (April 2026) is a 1T-parameter MoE with a 256K context window at $0.60/$2.80 — an aggressive-value agentic model. K3 nearly triples the parameter count to 2.8T, quadruples the context window to 1M, adds a deeper multimodal reasoning stack, and quintuples the price to $3/$15. That's a repositioning, not just a scale-up: K2.6 competed on value, K3 competes on capability. Both stay in Council's lineup — seat K2.6 when the loop is long but the task is contained, K3 when the repository is huge or the plan horizon is measured in hours.

How Council AI runs it

Kimi K3 is available on every paid Council AI plan (Plus, Pro, Ultra), routed through OpenRouter rather than a dedicated Moonshot integration — functionally identical from the user's side. In a council it runs in parallel with models from up to 8 other labs, and the moderator synthesizes a consensus answer with an agreement score. K3 earns its seat as the open-weight counterweight: when the closed Western flagships all agree, K3's independently-trained perspective is a genuinely uncorrelated check — and when it dissents on a coding question, it's often because it navigated the repo differently.

Frequently asked questions