Moonshot Kimi K3 reaches API users as older Kimi models sunset

Moonshot's Kimi K3 is now an API reality, not just a model announcement. Kimi's official docs list kimi-k3 as the flagship model for long-horizon coding, knowledge work, and deep reasoning, and the model list says older Kimi lines are closing to new users ahead of a full August 31 sunset.

That combination makes K3 a migration event. Developers get a new 2.8-trillion-parameter, native-vision model with a 1M-token context window, while teams still pinned to kimi-k2.5 or moonshot-v1 have a dated reason to move.

Kimi K3 is live in the API docs

The Kimi quickstart describes K3 as Moonshot's most capable flagship model to date. It uses Kimi Delta Attention and Attention Residuals, supports native visual understanding, and is positioned for frontier workloads such as long-horizon coding and end-to-end knowledge work.

The docs show standard OpenAI SDK-compatible examples against https://api.moonshot.ai/v1 with the model name kimi-k3. They also document K3-specific behavior: thinking mode is always enabled, reasoning_effort supports low, high, and max, and streaming can return separate reasoning and final-answer deltas.

For very large workflows, the important number is context. Kimi lists a 1M-token window and says context caching is automatic when long prefixes stay unchanged.

The older Kimi API path now has a deadline

Kimi's model list adds the operational pressure: after the K3 launch, kimi-k2.5 and the moonshot-v1 series are no longer available to newly registered users, with a full platform sunset on August 31, 2026.

That affects integrations that treated older model IDs as stable defaults. New accounts need to start on the current Kimi model line, and existing users should test K3 behavior before the cutoff. The changes are especially relevant for agents, coding assistants, and document workflows where reasoning traces, multimodal inputs, or context limits can alter application behavior.

There are also limits to account for. Kimi says vision inputs cannot use public image URLs; developers need base64 data or Kimi file references. The docs also say web search is being updated and is not recommended for production workflows in the near term.

Demand is already testing capacity

The Associated Press reported that Kimi K3 suspended new subscriptions after demand pushed capacity close to its limits. Treat that as an availability warning, not a benchmark conclusion. If your product depends on K3, test rate limits, fallback routes, and account eligibility before promising it to customers.

OpenRouter also lists Kimi K3, which gives teams another availability surface to watch. That does not replace vendor verification; it is useful for confirming model distribution and comparing routing options.

For LinkLoot readers, the practical move is simple: inventory every kimi-k2.5, moonshot-v1, and Kimi-compatible agent config, then run a side-by-side K3 test on real long-context and visual workloads. If you are building agent stacks, keep a fallback model in your runbook and compare it with the tools in LinkLoot's AI agent tools guide.

Evidence

The primary source is Kimi's own K3 quickstart and model list. AP provides independent context on demand and subscription pressure, while OpenRouter confirms third-party model availability. The full model weights are described by Kimi as scheduled for release by July 27, 2026, so this post treats open-weight publication as a pending milestone, not a completed release.