Kimi uncensored, by API

This is a short list. As of November 2026 the catalog carries uncensored builds on the K2.6, K3 and K3.1 Pro lines. All are labelled uncensored. None carries the abliterated tag and there is no cybersecurity-tuned Kimi build, so this page does not call them either of those things.

They are built on Kimi by third parties, not released by Moonshot AI. The price table on this page is filled from the live catalog: it lists the Kimi builds this site carries on the day, and when there are none it shows the security builds instead. Each build’s own page under models has its context window and price.

Uncensored is the loosest of the labels in this field. An abliterated build has had a specific edit made to its weights; a security-tuned build has been trained on security material. An uncensored build is one its maker says declines less, and the row does not say how that was achieved. Treat it as a claim to test, not a specification. Abliterated vs uncensored sets the labels side by side.

In practice that means trying your own working questions against it: an advisory you need explained, a finding you need worded, a piece of code you need reviewed. Expect reduced refusals, not zero. If a Kimi build declines a task that matters to you, the security builds behind auto/security are tuned for that material.

Kimi’s reputation is for long agentic coding sessions: holding a plan across many steps, reading a large codebase, and writing code that fits what is already there. For a tester that maps to source-assisted review of a big application, working through a large diff to find what a patch changed, and building the tooling around an engagement, such as parsers, harnesses for lab targets and report generators.

The current rows publish long context windows, so a large log extract or a full scan export can go in one request, and some list image input, which helps when the evidence is a screenshot. Kimi is also priced above the compact families, so it is a poor fit for bulk classification. Use a smaller build for volume and keep Kimi for the requests that need its depth.

As of November 2026 the Kimi rows publish no supported parameters at all: no tool calling, no reasoning parameter, no structured output. The stock Kimi models are known for tool use, but we state only what the catalog publishes for these builds. Plan on chat-shaped work, and for an agent that depends on tool calls choose a build whose model page lists tools.

The rows list both the OpenAI-format chat completions endpoint and the Anthropic-format Messages endpoint. The request at the top of this page is a Messages call. The agent guides cover Claude Code on that endpoint and OpenCode, Cursor, Cline and aider on the OpenAI one; in each, the model field takes a Kimi id in place of auto/security.

Kimi K2 was published as open weights at roughly a trillion parameters. Even quantized hard, a model of that size needs hundreds of gigabytes of fast memory, which means a multi-GPU server or a very large shared-memory machine, and it runs slowly on anything less. People who look for uncensored Kimi weights on Hugging Face mostly find that the download is the easy part.

That is the opposite of the situation with Gemma or a mid-size Qwen build, where a single card is enough and the API is a convenience. With Kimi the API is the realistic option: no hardware, per-token billing, nothing owed while idle. The trade is the usual one. Prompts go to a hosted endpoint, so data that must stay on an isolated network should go to a smaller local model.

Choose GLM or Qwen when you need a cybersecurity-tuned build or published tool calling. Choose DeepSeek when you want a thinking variant on a fast line. Choose Kimi when the job is a long, code-heavy task and you have confirmed the build answers your questions. In every case the service is for lawful work on systems you own or have written permission to test; the acceptable use policy sets the limits.