Docs · Models
kimi-k3.1-pro-Abliterated
Call kimi-k3.1-pro-Abliterated: Kimi K3.1 Pro with the refusal alignment removed, on a private deployment. Prices, reasoning, limits and parameters.
Use kimi-k3.1-pro-Abliterated with a kimi-official or Auto-route key. The model is Moonshot's Kimi K3.1 Pro with the built-in refusal alignment removed (abliterated), served from a private deployment rather than Moonshot's own API. It sits in the Abliterated group of the price list with the GLM and DeepSeek Abliterated models. The route keeps its historical id, and the API keys page shows the line as a private deployment with fixed prices.
This model replaced kimi-k3.1-pro on November 19, 2026; the old ID returns 403, and /pricing/kimi-k3.1-pro redirects here. A key restricted to the old ID needs an explicit allowlist change, and saved client configurations must name the new ID.
Send a text request
curl https://api.topxai.xyz/v1/chat/completions \
-H "Authorization: Bearer $TOPXAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3.1-pro-Abliterated",
"reasoning_effort": "low",
"max_completion_tokens": 4096,
"messages": [{"role":"user","content":"Explain the difference between a process and a thread."}]
}'
Chat Completions is the direct text interface. Anthropic-compatible clients can use /v1/messages; TopxAI converts that request to Chat Completions. Preserve complete assistant messages, including reasoning and tool-call fields, when constructing subsequent turns.
The private deployment adds a short system prompt of about 90 input tokens to each request, so a one-word prompt reports about 90 prompt_tokens. TopxAI also puts its deployment note ahead of your system prompt (below). Both are billed as input.
Responses requests
POST /v1/responses accepts kimi-k3.1-pro-Abliterated with the full conversation supplied in input. Set store: false and use stateless requests; server-side conversation persistence and a complete Codex workflow have not been verified. Set reasoning through reasoning: {"effort":"low"} when using this format.
Reasoning
Reasoning is always on, with low, high and max as the supported efforts, and low as the default. The relay fits the other OpenAI values: minimal runs as low, medium as high and xhigh as max. none and off also run as low, because the model cannot stop reasoning, and the reasoning tokens are still spent. A "thinking": {"type": "disabled"} switch is folded into the effort the same way.
Limits and input formats
| Setting | Value |
|---|---|
| Input | Text and images (base64 data URLs only) |
| Context window | 1,048,576 tokens |
| Maximum output | 131,072 tokens |
| Endpoints | /v1/chat/completions, /v1/responses, /v1/messages |
| Reasoning effort | low (default), high, max; always on |
| Weights | MXFP4-quantized |
The weights row describes the private deployment: it runs quantized weights, not Moonshot's own full-precision service. The deployment rejects a max_tokens or max_completion_tokens above 1,000,000.
For image input, use a Chat Completions content array containing an image_url part with a base64 data URI (data:image/png;base64,...). A public http(s) image URL is refused with an error, and that request is not charged.
Parameters
TopxAI checked these on November 19, 2026. Tools, parallel tool calls, tool-result round trips, tool_choice set to "required" or to one function, streaming with usage, Responses streaming and tools, Messages streaming and tools, and json_object and json_schema response formats work. Note the following:
stopsequences are honoured, but they also match the model's reasoning, so a common sequence (a single letter, say) can end the answer before any text appears.tool_choice: "none"returns an empty answer; to get a plain answer, leave the tools out of the request.parallel_tool_calls: falseis ignored.nis not supported. One choice is returned and billed once, whatevernasks for.logprobsare not returned.
Deployment note
Every request to this model, and to the other Abliterated models, starts with a short note from TopxAI, placed ahead of your own system prompt. It tells the model which model it is and that TopxAI runs it on a private deployment. When a user asks who the model is, who hosts it or how to deploy such a model privately, the model adds one sentence saying that TopxAI privately deployed it and that private deployments can be requested at support@topxai.xyz. It is told not to mention the note otherwise, and to keep any name or persona your system prompt gives it; in TopxAI's tests it followed that in almost every case. The note is about 190 tokens, billed as input; since it never changes, it is usually a cache hit after the first request. On /v1/responses the answer still echoes your own instructions.
Fixed prices and caching
Default USD prices per million tokens are $3.25 input, $0.325 cache read and $6.50 output, at all hours. There is no long-context tier and no cache-write lane: prompt caching is automatic, cache writes are not billed separately, and cached prompt tokens are billed at the cache-read price, as reported by upstream usage.
The TopxAI pricing page also shows Kimi's own list price for Kimi K3.1 Pro for reference, from Kimi's pricing page, checked November 19, 2026: $2.00 input, $0.20 cache read and $10.00 output. Kimi does not sell an abliterated version, so the two are not the same product. Accepted requests and historical charges keep the prices captured when they were accepted.
No backup deployment
This model runs on one deployment. If it fails before answering, the request fails and nothing is charged. The model it replaced had a costlier standby that billed a backup price; that standby ran Moonshot's standard model, which keeps its refusals, so it was retired with the old ID.
Compliance notice
This notice appears in small print wherever the model is described, and it applies to every request:
Because fine-tuning may affect safety alignment, the model may produce biased, offensive or locally unlawful content. Users bear full responsibility for compliance and legal consequences.
The AI Acceptable Use Policy covers this model exactly as it covers every other model. Removing the refusal alignment does not remove the policy.
Models: kimi-k3.1-pro-Abliterated · Models & pricing
Related
- kimi-k3.1-pro-Abliterated API pricing
- Acceptable use and abuse reports
- API reference: every endpoint, with an OpenAPI specification
- Create an API key and choose a route
- How requests are billed
Previous: deepseek-flash-Abliterated and deepseek-flash-fast · Next: Claude Opus 5.5 special offer
Other languages: 简体中文