Kimi K3.1 Pro
Abliterated API.
Moonshot AI’s flagship model family on the same Oathbreaker key and prepaid balance. Text chat, streaming and function calling through an OpenAI-compatible API.
Same account. A larger model.
Oathbreaker Kimi gives developers access to Kimi K3.1 Pro Abliterated, built from Moonshot AI’s Kimi K3.1 Pro, through an OpenAI-compatible Chat Completions API. It supports text conversations, streaming responses, and function calling for agents that run tools in your own environment.
Oathbreaker is operated by Formacha LLC. We provide the API gateway, account, keys, and usage accounting. The model family comes from Moonshot AI; an inference provider supplies the modified checkpoint. Oathbreaker does not train the underlying model.
Set model to oathbreaker-kimi-k3.1-pro and use https://oathbreaker.ai/v1 as your base URL. Kimi is an explicit choice; it does not replace the flagship GLM-5.3 Abliterated default.
Kimi pricing
Pay for confirmed token usage from your prepaid balance. Kimi uses the same account and API key as the other models, and costs more per token than they do.
| Token type | Price |
|---|---|
| Input | $4.50 |
| Cached input | $0.45 |
| Output, including reasoning | $9.00 |
For example, 100,000 uncached input tokens plus 10,000 output tokens cost $0.54. This is a token-cost illustration, not a benchmark or a promise about how many tokens an agent task will use.
Kimi has one price for everyone: monthly plan members pay the same rate, while member rates still apply on the other models. Add credit starting at $5. Cached input is charged once at its discounted rate when the provider reports a cache hit; repeated prompts do not guarantee one. See billing details and pricing for every model.
Capabilities & current limits
These are the limits exposed by Oathbreaker. The base model’s larger advertised context window and image support do not change this API’s limits.
| Feature | Oathbreaker Kimi |
|---|---|
| Input / output | Text / text |
| Admission budget | 131,072 units shared by UTF-8 input bytes, message and tool framing, and requested output tokens |
| Maximum output | 24,576 tokens per request, including reasoning; larger budgets are lowered to this; 8,192 if you do not set a budget |
| Request size / history | 262,144 bytes; up to 2,048 messages |
| Streaming | Server-sent events (SSE), including final usage |
| Function calling | Up to 128 tool definitions; your application executes the tools |
| Reasoning | Low, high, or max effort; reasoning shares the output budget |
The admission budget is conservative: it counts bytes rather than tokens, so a request can reach it before its token count does. It is not a tokenizer-measured context window.
Kimi reasons before it answers, and reasoning tokens are billed as output. If you omit reasoning_effort, Oathbreaker sends low. Set high or max when a task needs more thinking, and raise max_tokens to leave room for the answer.
This is a text API: image input, audio and embeddings are not supported. JSON response formats are passed to the model; validate returned JSON and tool arguments in your application. Account rate limits are shared across models. Read the full Kimi limits and rate limits.
Your first Kimi request
Create an account, add prepaid credit, and create an API key in your console. Store it as OATHBREAKER_API_KEY in your terminal or server environment. Running these examples uses credit.
curl --fail-with-body --silent --show-error --max-time 650 \
https://oathbreaker.ai/v1/chat/completions \
-H "Authorization: Bearer $OATHBREAKER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "oathbreaker-kimi-k3.1-pro",
"messages": [{"role": "user", "content": "Suggest three tests for a CSV import tool."}],
"max_tokens": 2048,
"stream": false
}'Using the OpenAI Python SDK
Install the openai package in your Python environment, then run:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://oathbreaker.ai/v1",
api_key=os.environ["OATHBREAKER_API_KEY"],
timeout=650.0,
max_retries=0,
)
response = client.chat.completions.create(
model="oathbreaker-kimi-k3.1-pro",
messages=[{"role": "user", "content": "Suggest three tests for a CSV import tool."}],
max_tokens=2048,
)
print(response.choices[0].message.content)
print(response.usage)These examples send one request and do not retry automatically. The output budget includes reasoning, so it may be used up before a final answer. Check the response’s finish_reason and your console usage before repeating a request.
For streaming, set stream: true and consume the SSE events. For tools, supply function definitions, validate returned arguments, and send the tool results back with the conversation. Follow the streaming guide and tool-calling guide for the complete request flow.
Connect Kimi to Hermes or another agent
Follow the Hermes setup guide using your Oathbreaker key and https://oathbreaker.ai/v1. When selecting a model, choose oathbreaker-kimi-k3.1-pro. Start a new conversation so the new model selection takes effect.
For OpenCode, OpenWork, or Pi, use the other agent setup guides with the same base URL and the Kimi model identifier. Use a custom OpenAI-compatible connection where required.
Start with a small, reviewable task in a folder you choose. Agents resend the conversation on every step, so a long session costs more on Kimi than on the other models. You can inspect requests and token costs in your console’s Usage tab.
Common questions
Read the abliterated-model guide for a fuller explanation, hosted-versus-local tradeoffs, and how to evaluate a model on your own task.
What does “abliterated” mean?
It describes a model modified to reduce learned refusal behavior. These models are sometimes called uncensored. That does not guarantee an answer to every prompt, correct output, or suitability for every task. Human judgment and responsibility still matter.
Where does the model come from?
Oathbreaker Kimi serves a hosted abliterated variant of Kimi K3.1 Pro. Moonshot AI retains attribution for the underlying model. Hosted checkpoints and performance can differ from the original release, and the inference provider currently lists this checkpoint as a preview.
Should I use Kimi, DeepSeek or GLM?
All use the same key and balance, so you can switch per request. Kimi is the largest model here and the most expensive per token. Send the same small task with a checkable result to each model, such as a code change with tests, and compare the answers, response time and token cost. See the GLM-5.3 Abliterated, GLM-5.3 Flash Abliterated and DeepSeek V4.1 Flash Abliterated pages.
Who can help with setup?
Email support@oathbreaker.ai with the error message and request ID, without your API key. The API reference, privacy policy, and terms cover request behavior and service policies.