Refuseless: hosted API for abliterated GLM models
A new hosted inference API called Refuseless puts abliterated open-weight models within reach of any application that can speak the OpenAI SDK format. The service runs Z AI's GLM-5.3 and GLM-5.3-Flash with the refusal direction removed from the weights. Point any OpenAI SDK at https://api.refuseless.com/v1, set the model to glm-5.3-abliterated and pay per million tokens. Prompts are not retained after the response so the service carries no conversation memory beyond a single request-response cycle.
What abliteration means in practice
Abliteration is the process of removing the refusal direction from a model's weights. The source models GLM-5.3 and GLM-5.3-Flash from Z AI include a built-in refusal direction that causes the model to decline requests that fall into certain content categories. Ablation strips that directional component so the model answers prompts it would otherwise refuse. The weights themselves are modified; no post-processing or prompt-engineering trick is needed. The abliterated model is distributed as part of the Refuseless service and charged like any other hosted LLM.
Connecting and using the service
Getting started requires three steps. First, obtain an API key from the Refuseless keys page. Keys start with the prefix abl_. Second, add the provider in OpenCode or any compatible tooling. In OpenCode the command is /connect, choose Refuseless and paste the key. If Refuseless does not appear in the built-in list, pick Other and enter refuseless as the provider id. The key is saved separately and takes priority over any apiKey in the config so a newly created key will not take effect until /connect is run again. Third, select the model. The full lineup of five models is already in the config so any can be switched to at any time. Running opencode --model refuseless/glm-5.3-abliterated or using /models in OpenCode and picking Refuseless › GLM 5.3 Abliterated will confirm the connection with a simple "Say ok." prompt.
Running requests against the API
The request shape is identical to OpenAI chat completions. The base URL is set to https://api.refuseless.com/v1 the model name is set to glm-5.3-abliterated (or any of the other four lineup models) and the API key is passed as authentication. Tokens are counted and charged per million. Because prompts are not retained after the response, each request is a standalone call. This makes the service suitable for stateless workloads, classification tasks, or any pipeline where the application manages conversation state externally.
Cost and model lineup
Pricing is per million tokens. The exact rate card is published on the Refuseless keys page alongside the API key. Five models are available in the config, ranging from the full GLM-5.3-abliterated to a Flash variant optimized for lower latency and cost. Developers can switch between models at any time because all five are pre-configured. The per-token cost model means you only pay for the tokens you consume, with no minimum monthly spend or reservation commitment.
Why this matters for tooling and workflows
Most hosted LLM APIs enforce refusal behavior as a fixed policy. Abliterated models give developers the option to accept or reject requests based on their own content policies rather than the provider's. This is relevant for applications that need to process user-generated content, run automated classification pipelines, or build assistants that must operate across jurisdictions with different content rules. Because the service is OpenAI-compatible, existing tooling works without modification. SDKs that already handle authentication, token counting and request formatting can point at Refuseless with a single configuration change.
Technical details and limitations
The abliteration is applied at the weight level so the model's core capabilities reasoning, code generation, function calling remain intact. Only the refusal direction is removed. This means the model may still decline requests for reasons unrelated to the removed direction, such as safety filters built into the inference pipeline. The service does not retain prompts beyond the immediate response so any session state must be managed by the calling application. Rate limits, timeout behavior and error formats follow the OpenAI convention so existing retry and backoff strategies apply without customization.
Getting access
Developers can sign up for a free key on the Refuseless website. The key is tied to an API keys page where usage and remaining balance are displayed. Once the key is obtained it is added through the /connect command or equivalent tooling and the desired model is selected from the five available. The service is already live and accepting requests.