Uncensored models through LiteLLM

Point LiteLLM at an uncensored OpenAI-compatible endpoint: openai/ prefix, api_base and key in the Python SDK and the proxy config.yaml, with a fallback chain.

Steps checked October 25, 2026 · prices as of November 16, 2026

Three settings connect LiteLLM to an OpenAI-compatible endpoint: a model string of openai/ plus an id, an api_base and a key. Below, those settings carry uncensored builds, first through the Python SDK and then through the proxy's config.yaml, ending in a fallback chain over three builds. Every field name was taken from docs.litellm.ai on 25 October 2026.

New to builds that had their refusals removed? Read abliterated vs uncensored first, then return here for the wiring.

What you need

model
openai/glm-5.3-uncensored
the openai/ prefix in front of an id from the models page
api_base
https://api.unrestrictable.ai/v1
stops at /v1
api_key
sk-unr-live-…
made in the console; exported as UNR_API_KEY

Make the key in the console. It begins with sk-unr-, is displayed one time, and draws on a prepaid balance you top up from $10. Every id is on the uncensored model list. To talk to the same builds without writing code, open /chat or message the Telegram bot.

Why the model name starts with openai/

LiteLLM reads the provider off the front of the model string. A server that implements OpenAI's /chat/completions gets openai/, then the id that server knows. The OpenAI-compatible endpoints doc explains the prefix in a single code comment: it tells litellm to route the call as OpenAI.

That doc also says to pass the base URL as is, ending in /v1. All the uncensored builds here answer in the OpenAI wire format, so one prefix serves the whole catalog. Checked against litellm 1.106.3, the PyPI release current on 25 October 2026.

First reply from the Python SDK

A single call with no config file. The result has the OpenAI shape, so anything written against the OpenAI SDK parses it unchanged.

main.py
1import os
2import litellm
3
4response = litellm.completion(
5 model="openai/glm-5.3-uncensored",
6 api_base="https://api.unrestrictable.ai/v1",
7 api_key=os.environ["UNR_API_KEY"],
8 messages=[{"role": "user", "content": "Outline a heist thriller in five beats."}],
9 max_tokens=600,
10)
11print(response.choices[0].message.content)

To try another build, replace the id behind openai/. The base URL and the key stay the same for every build in the catalog.

Three builds in one fallback chain

A Router holds named deployments plus a fallbacks map. The reliability doc describes the behaviour: once a call has used up num_retries, LiteLLM tries the next name in the list, in the order written.

router.py
1import os
2from litellm import Router
3
4BASE = "https://api.unrestrictable.ai/v1"
5KEY = os.environ["UNR_API_KEY"]
6
7
8def build(model_id):
9 return {"model": f"openai/{model_id}", "api_base": BASE, "api_key": KEY}
10
11
12router = Router(
13 model_list=[
14 {"model_name": "uncensored-primary", "litellm_params": build("glm-5.3-uncensored")},
15 {"model_name": "uncensored-second", "litellm_params": build("deepseek-v4-pro-uncensored")},
16 {"model_name": "uncensored-third", "litellm_params": build("glm-5.3-abliterated")},
17 ],
18 fallbacks=[{"uncensored-primary": ["uncensored-second", "uncensored-third"]}],
19 num_retries=1,
20)
21
22response = router.completion(
23 model="uncensored-primary",
24 messages=[{"role": "user", "content": "Outline a heist thriller in five beats."}],
25)
26print(response.choices[0].message.content)

One key sits behind all three entries, so a 401 or a 402 returns identically from each; a fallback does not route around it. The chain earns its keep when one build alone fails: a 404 on an id that has left the catalog, a 429, a 5xx, a timeout. As the docs page says, a request that dies before the model answers is never billed, so a request that falls through pays once, at the rate of the build that answered.

The names are aliases. The application calls uncensored-primary and never learns the model id, so swapping builds is one line in the list. num_retries=1 grants a failing build a second attempt before the chain advances; a retry that fails before the model answers is likewise unbilled.

The chain as a proxy config.yaml

The proxy loads the same deployments from a file and exposes them on its own OpenAI-format endpoint. os.environ/ pulls the key from the environment, so it never lands in the file. The numbers under model_info are this catalog's per-token prices, filled in when this page rendered, so LiteLLM's spend tracking lines up with the API's bill.

config.yaml
1model_list:
2 - model_name: uncensored-primary
3 litellm_params:
4 model: openai/glm-5.3-uncensored
5 api_base: https://api.unrestrictable.ai/v1
6 api_key: os.environ/UNR_API_KEY
7 model_info:
8 input_cost_per_token: 0.000000282
9 output_cost_per_token: 0.000000846
10 - model_name: uncensored-second
11 litellm_params:
12 model: openai/deepseek-v4-pro-uncensored
13 api_base: https://api.unrestrictable.ai/v1
14 api_key: os.environ/UNR_API_KEY
15 model_info:
16 input_cost_per_token: 0.00000097
17 output_cost_per_token: 0.00000194
18 - model_name: uncensored-third
19 litellm_params:
20 model: openai/glm-5.3-abliterated
21 api_base: https://api.unrestrictable.ai/v1
22 api_key: os.environ/UNR_API_KEY
23 model_info:
24 input_cost_per_token: 0.0000021825
25 output_cost_per_token: 0.000004365
26
27router_settings:
28 fallbacks: [{"uncensored-primary": ["uncensored-second", "uncensored-third"]}]
29
30litellm_settings:
31 num_retries: 1
32 request_timeout: 120
33
34general_settings:
35 master_key: os.environ/LITELLM_MASTER_KEY
shell
1export UNR_API_KEY="sk-unr-live-…"
2export LITELLM_MASTER_KEY="sk-pick-your-own"
3litellm --config config.yaml

The proxy comes up on port 4000. Give any OpenAI SDK that address and the master key, and request uncensored-primary as the model. In LiteLLM's custom pricing doc the two *_cost_per_token keys are for tracking; they do not change what is billed.

Fallbacks for context size and content filters

context_window_fallbacks catch LiteLLM's ContextWindowExceededError. The reliability doc requires enable_pre_call_checks: true in router_settings for that check; without it, requests of any size go straight to the provider. The check also needs a published window, and the three builds in the chain publish none in the catalog. For a context fallback, add a fourth deployment on nemotron-3-ultra-550b-a55b-uncensored, whose catalog entry lists 262,144 tokens, and point uncensored-primary at it.

content_policy_fallbacks respond to a provider's content-filter error, LiteLLM's ContentPolicyViolationError. Map a filtered model elsewhere to uncensored-primary and those requests land here. A refusal typed into an ordinary reply arrives as a success, not an error, so no fallback can catch it.

Picking the three builds

Lead with the build you prefer; the backups only bill for the requests that reach them. At the time of writing, the three below had held steady since 12 October on the daily real-model check, which posts one fixed prompt to every model each day and records whether the reply still matches.

Uncensored models and their prices per million tokens
Wire formats
GLM-5.3 Uncensoredglm-5.3-uncensoredUncensored$0.282$0.846—openai
DeepSeek V4 Pro Uncensoreddeepseek-v4-pro-uncensoredUncensored$0.97$1.94—openai
GLM-5.3 Abliteratedglm-5.3-abliteratedAbliterated$2.1825$4.365—openai

Prices per million tokens, read from the live catalog on 2026-11-16.

Model idName in config.yamlToday's checkContext window
glm-5.3-uncensoreduncensored-primarySteady since Oct 12not published
deepseek-v4-pro-uncensoreduncensored-secondSteady since Oct 12not published
glm-5.3-abliterateduncensored-thirdSteady since Oct 30not published

Context windows from the catalog on 2026-11-16; check verdicts read from /proof when the page rendered.

With two GLM builds and one DeepSeek build, an outage on either family leaves a working route. More builds from each family are on the uncensored GLM and uncensored DeepSeek pages.

The bill for a batch

Take 10,000 requests, each 1,200 prompt tokens in and 350 reply tokens out, with one build answering all of them.

Model10,000 requests
GLM-5.3 Uncensored$6.35
DeepSeek V4 Pro Uncensored$18.43
GLM-5.3 Abliterated$41.47

Worked out at render from the catalog on 2026-11-16: requests × 1,200 × the input rate, added to requests × 350 × the output rate.

In practice a month lands near the first row, plus whatever the backups charge on the requests that reached them. The full rate card is the pricing page.

Test the key without LiteLLM

One request returns the ids the key is allowed to call. Nothing is charged for it, so run it before you suspect the config.

curl
1curl https://api.unrestrictable.ai/v1/models \
2 -H "Authorization: Bearer sk-unr-live-…"

A 200 carrying a list of ids proves both the key and the base URL. A 401 tells you how the key went wrong:

http
1HTTP 401
2{"error":{"message":"No API key was sent.","type":"authentication_error","param":null,"code":"invalid_api_key"}}
http
1HTTP 401
2{"error":{"message":"The API key is not valid.","type":"authentication_error","param":null,"code":"invalid_api_key"}}

Response bodies captured from the API on 25 October 2026.

LiteLLM errors and what they mean here

Per LiteLLM's exception mapping doc, each HTTP status becomes a class: AuthenticationError for 401, PaymentRequiredError for 402, NotFoundError for 404 and RateLimitError for 429.

Every call returns Not Found

Follow the OpenAI-compatible doc: api_base has to end in /v1. Then confirm the model string begins with openai/.

One build returns 404 model_not_found

The id behind openai/ is not in the catalog. The chain carries on to the next deployment; correct the id against the list the curl above returns.

http
1HTTP 404
2{"error":{"message":"No such model: glm-5.3-uncensord.","type":"not_found_error","param":null,"code":"model_not_found"}}

402 insufficient_balance

The prepaid balance has run out, and every build in the chain reports the same. Top up in the console.

http
1HTTP 402
2{"error":{"message":"The prepaid balance does not cover this request. Top up at … to continue.","type":"insufficient_quota","param":null,"code":"insufficient_balance"}}

Questions

In both directions. A model string that begins with openai/ sends the call to an OpenAI-compatible endpoint, and the proxy itself serves an OpenAI-format endpoint on port 4000.

A server that accepts and returns the same bodies as OpenAI's /v1/chat/completions, so a client needs only a base URL, a key and a model id. Here the base URL is https://api.unrestrictable.ai/v1. See the docs page.

Three values: model as openai/ plus the provider's id, api_base as its /v1 URL, and api_key as its key. For this API that means openai/glm-5.3-uncensored with https://api.unrestrictable.ai/v1.

Any service that answers the OpenAI wire format from its own base URL. LiteLLM’s doc lists dedicated prefixes too, hosted_vllm among them, for servers that fit one; this API uses the plain openai/ prefix. See the uncensored model list.

Yes. Because the proxy answers in the OpenAI format, the OpenAI Node SDK connects with baseURL pointed at the proxy and the master key as apiKey.

Only when it has the prices. The config.yaml above copies this catalog’s per-token rates into model_info; LiteLLM uses them for tracking and nothing else. See per-token prices.

Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Join on Telegram