Refuseless

Use Refuseless through LiteLLM with the refuseless/ model prefix. LiteLLM reads REFUSELESS_API_KEY and sends requests to https://api.refuseless.com/v1.

Get started

Create an API key in your Refuseless account and set REFUSELESS_API_KEY in your environment or secret manager. Install litellm for the Python SDK, or litellm[proxy] for the gateway.

export REFUSELESS_API_KEY="YOUR_REFUSELESS_API_KEY"
pip install litellm

from litellm import completion

response = completion(
    model="refuseless/glm-5.3-abliterated",
    messages=[{"role": "user", "content": "Reply with hello."}],
    reasoning_effort="low",
)
print(response.choices[0].message.content)

A successful request returns an assistant reply and token usage. You do not need to set api_base.

Models and pricing

Prices below are USD per million tokens, checked against the published pricing and live model catalog on October 26, 2026. Context is the combined input and output budget, so the full output limit is available only when the prompt fits in the remaining context. Prefix each model ID with refuseless/ in LiteLLM.

Model ID	Context	Max output	Input	Cached input	Output
glm-5.3-abliterated	1,048,576	131,072	$1.25	$0.25	$2.25
glm-5.3-flash-abliterated	1,048,576	131,072	$0.25	$0.15	$0.8
deepseek-v4.1-flash-abliterated	1,048,576	130,000	$0.22	$0.045	$0.75
deepseek-v4.1-flash-fast	1,048,576	262,144	$0.05	$0.003	$0.15
minimax-m3.1-flash	1,048,576	512,000	$0.05	$0.002	$0.25

The first three models use abliterated weights. The other two use unmodified weights.

Refuseless also lists deepseek-v4.1-flash-free, but live inference returned 404 model_not_found at verification. LiteLLM excludes it from its model catalog and the supported-model table above until inference works. Choose one of the five verified models instead.

Gateway setup

Install the proxy with pip install 'litellm[proxy]'. Set REFUSELESS_API_KEY on the gateway process, and set LITELLM_MASTER_KEY to a strong key for gateway clients. The provider key authenticates upstream requests. The gateway key authenticates your client to LiteLLM.

Add this deployment to your existing configuration, or save it as config.yaml for a new gateway.

model_list:
  - model_name: refuseless-chat
    litellm_params:
      model: refuseless/glm-5.3-abliterated
      api_key: os.environ/REFUSELESS_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

litellm --config config.yaml --port 4000

In a second terminal with LITELLM_MASTER_KEY set, send a request.

curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "refuseless-chat",
    "messages": [{"role": "user", "content": "Reply with hello."}],
    "reasoning_effort": "low"
  }'

Expect HTTP 200 with an assistant message in choices[0].message.content and counts in usage.

Streaming and prompt caching

Set stream=True and request usage in the final chunk.

from litellm import completion

stream = completion(
    model="refuseless/glm-5.3-abliterated",
    messages=[{"role": "user", "content": "Reply with hello."}],
    reasoning_effort="low",
    stream=True,
    stream_options={"include_usage": True},
)
for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="")
    if getattr(chunk, "usage", None):
        print(chunk.usage)

LiteLLM uses usage.prompt_tokens_details.cached_tokens to charge reused input at the cached-input rate. Uncached input uses the standard input rate. A repeated prompt does not guarantee a cache hit.

Tools, structured output, and reasoning

Chat Completions supports tools, tool_choice, and response_format. Use LiteLLM function calling for tool-result continuations and structured outputs for JSON Schema examples.

Use reasoning_effort="low", "high", or "max". The live catalog publishes supported effort levels per model and defaults to low. Refuseless also advertises off for its paid DeepSeek variants. Do not assume OpenAI effort names such as medium and xhigh are accepted by the provider. LiteLLM uses the declared effort levels to resolve supported requests. Reasoning text, when returned, is available as message.reasoning_content. See Refuseless reasoning.

Responses API

LiteLLM routes responses() to the provider's native /v1/responses endpoint. Send conversation history in input. These APIs are stateless, so do not rely on previous_response_id for stored history.

from litellm import responses

response = responses(
    model="refuseless/glm-5.3-abliterated",
    input="Reply with hello.",
    reasoning={"effort": "low"},
    store=False,
)
print(response.output)

Gateway clients can send the same request to /v1/responses with the configured deployment name.

Troubleshooting

For 401, check the key in the gateway process environment. For 402, check provider credits. For 404, compare the unprefixed model ID with GET https://api.refuseless.com/v1/models. For 429, honor the provider's retry guidance. A working model catalog does not guarantee that every listed model is currently available for inference.