Abliterated models, without a GPU

Set three values

To use an abliterated model without a GPU, configure an OpenAI-compatible client with the base URL https://uncensuredai.com/v1, an Uncensured AI key and the chosen variant’s model ID. Inference runs remotely; your computer does not load the model weights.

The Uncensured AI API documentation specifies Chat Completions and Bearer authentication. This guide concerns Uncensured AI at uncensuredai.com, rather than similarly named services.

The examples are documented configurations, not authenticated calls executed by Deepweight. We checked the public endpoint documentation and catalogue; an end-to-end test requires a valid API key and available credit.

Prepare your client

Create a key in the service console and keep it on your backend. For Python, install the official openai package in a virtual environment. Supply UNCENSURED_API_KEY to the process through your environment or secret manager, rather than committing it to source control.

python -m venv .venv
source .venv/bin/activate
python -m pip install openai

For a web application, send the request through a server route you control. A key embedded in browser JavaScript can be read by visitors. Separate application keys make it possible to revoke one integration without interrupting the others.

Make a first request

This configuration uses the public ID for DeepSeek V4.1 Flash Abliterated. An explicit timeout and disabled automatic retries keep the first experiment easier to inspect. Add a retry policy after verifying the basic request path.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["UNCENSURED_API_KEY"],
    base_url="https://uncensuredai.com/v1",
    timeout=60.0,
    max_retries=0,
)

response = client.chat.completions.create(
    model="deepseek-v4.1-flash-abliterated",
    messages=[{
        "role": "user",
        "content": "Explain an LRU cache with a Python example.",
    }],
    max_tokens=800,
)

print(response.choices[0].message.content)
print(response.usage)

The SDK handles transport; protocol compatibility does not make every feature identical. Before enabling structured output, images, tool calls or reasoning controls, check support for the specific served variant. Do not move provider-specific options between endpoints without testing them.

Choose the right variant

The model IDs checked in the public catalogue are deepseek-v4.1-flash-abliterated, glm-5.3-flash-abliterated and glm-5.3-abliterated. Changing the ID changes the served model and its price. An -uncensored suffix identifies a different edition, not a spelling variant.

For a short development loop, start with a Flash model and limit output to a few hundred tokens. Keep a small set of representative tasks: JSON transformation, a small bug with tests, a summary and a question with missing context. This is not a scientific benchmark, but it is better than choosing from one lucky answer.

Compare the cost of accepted results. Our guide to choosing abliterated model APIs includes verified prices and a consistent volume example.

Add streaming

In a chat, showing early fragments reduces the wait for the complete answer. A network fragment is not a model token: an event can contain several tokens or only metadata.

stream = client.chat.completions.create(
    model="deepseek-v4.1-flash-abliterated",
    messages=[{"role": "user", "content": "Describe an LRU cache."}],
    max_tokens=800,
    stream=True,
)

try:
    for chunk in stream:
        if chunk.choices:
            text = chunk.choices[0].delta.content
            if text:
                print(text, end="", flush=True)
finally:
    stream.close()

In the browser, provide a Stop button and preserve the text already received. The guide to SSE streaming without losing text covers network boundaries, UTF-8 decoding and event separators.

A migration still needs a test

With an OpenAI-compatible client, the transport change is small: key, base URL and model name. Functional verification is broader. Check error handling, timeouts, response shape, usage, multi-turn messages and tools. Begin with low concurrency and route a small share of traffic to the new endpoint after comparing outputs.

The figures communicated to the publisher for Uncensured AI are 250–300 tokens/s on average, depending on the model, and a 0% refusal rate. These are not independent Deepweight measurements and do not describe time to first token. For capacity planning, record total duration, billed tokens and accepted results: isolated throughput is not enough to size a queue.

An API avoids buying hardware, but it remains a usage-based expense. Set output limits, request limits and a spending budget. For 429 or 503 errors, use jittered backoff within a total time limit. Avoid stacking a retry wrapper on top of SDK retries without counting the combined attempts.

Before going live

Can I use Cline or OpenCode?

The provider publishes a Cline guide and an OpenCode guide. Follow the documented OpenAI-compatible configuration for your app version.

Is changing one base URL enough?

It changes the destination, not the proof of equivalence. You also need a key for the new service, an existing model ID and supported parameters.

Are the tests independent?

A test labelled as a Deepweight test must be run and documented by the editorial team, separately from provider figures. This guide presents no executed API benchmark. Its communicated performance figures are identified as such; the methodology page explains what reproducible measurements require.