# Abliterated Models: How to Call Them Through an OpenAI-Compatible API By Tim · 20 Nov 2026 · Launch Spotlights > What abliterated models are, what to check in a hosted provider, and how to call UncensoredLabs' abliterated DeepSeek and GLM models from an existing OpenAI client, with pricing and data handling. --- # Abliterated Models: How to Call Them Through an OpenAI-Compatible API An abliterated model is a language model with its refusals removed from the weights. You can run the weights yourself, or call a hosted endpoint and pay only for the tokens you use. This guide covers the hosted route: what to check in a provider, and how to point an existing OpenAI client at one. [UncensoredLabs](https://uncensoredlabs.dev/) is one such provider. It serves abliterated and uncensored variants of DeepSeek and GLM, alongside unmodified standard models from DeepSeek, MiniMax and Kimi, through browser chat and an OpenAI-compatible API at `https://uncensoredlabs.dev/v1`. Billing is prepaid and per token, with top-ups from $3. Below is how the catalog, setup, billing and data handling work, based on UncensoredLabs' own site and docs. ![UncensoredLabs homepage with a sample console showing per-request spend for abliterated DeepSeek and GLM models](https://res.cloudinary.com/dhdy0tbpt/image/upload/c_crop,w_1440,h_1400,x_0,y_0/v1794677302/ShipBoost/screenshots/screenshot-51186fff-0f39-48cc-ae37-a86e00eb57ad.webp) ## What the labels mean in the catalog UncensoredLabs sorts every model into one of three classes, and the class is a field in the API's model list: - **Standard:** the original, unmodified weights. The catalog describes these as "original unmodified" DeepSeek, MiniMax and Kimi weights. - **Abliterated:** the catalog's own description is "refusals removed from the weights". - **Uncensored:** a separate class, applied to one DeepSeek and one GLM model. The site doesn't describe how these variants differ from the abliterated ones, so check the [model pages](https://uncensoredlabs.dev/models) before you choose. The catalog is the source of truth for which model IDs are enabled, what each one supports, and what it costs. According to the docs, publishing a model is an operator action that requires verified identity, capabilities, prices and resale permission. ## What to check before you pick a hosted provider These questions decide whether a hosted provider fits your project: | What to check | Why it matters | UncensoredLabs | |---|---|---| | API shape | Whether your existing client works unchanged | OpenAI Chat Completions, custom base URL | | Streaming and tool calling | Agents and chat UIs need both | Streaming on all models; tool calling on models whose catalog entry lists it | | Context and output limits | Long documents and long generations | 1,048,576-token context on every listed model; max output 65,536 to 512,000 tokens depending on the model | | Billing model | Whether you can test cheaply | Prepaid credit from $3, pay per token, credits don't expire | | Prompt retention | Who keeps your inputs | Application servers don't store prompts or responses; the inference provider's retention policy applies | | Spend controls | Stops a runaway script draining your balance | Per-key spending limits and model allowlists | ## The UncensoredLabs model catalog At the time of writing, the public model list at `https://uncensoredlabs.dev/v1/models` returns nine models. Prices are per million tokens. | Model ID | Class | Max output (tokens) | Input / 1M | Output / 1M | |---|---|---|---|---| | `deepseek-v4.1-flash-1017-abliterated` | Abliterated | 131,072 | $0.12 | $0.45 | | `deepseek-v4.1-flash-abliterated-fast` | Abliterated | 65,536 | $0.12 | $0.45 | | `glm-5.3-abliterated` | Abliterated | 131,072 | $1.00 | $2.00 | | `glm-5.3-flash-abliterated` | Abliterated | 131,072 | $0.12 | $0.50 | | `deepseek-v4-pro-uncensored` | Uncensored | 65,536 | $0.50 | $0.50 | | `glm-5.3-uncensored` | Uncensored | 65,536 | $0.35 | $0.35 | | `deepseek-v4.1-flash-fast` | Standard | 262,144 | $0.01 | $0.08 | | `minimax-m3.1-flash` | Standard | 512,000 | $0.02 | $0.08 | | `kimi-k3.1-pro` | Standard | 131,072 | $0.25 | $2.50 | Cached input is billed at a separate, lower rate for each model. Live prices, capabilities and context limits are on the [models page](https://uncensoredlabs.dev/models). The homepage's model cards also show average tokens per second and a 24-hour success rate, and UncensoredLabs says measurements it doesn't have stay marked as unknown. ## How to connect an existing OpenAI client The [quickstart](https://uncensoredlabs.dev/docs/quickstart) follows the same loop the docs describe as the core workflow: sign in, check credit, select a model, send a request, review its usage. 1. **Create an account.** UncensoredLabs accounts can be created with an account key alone, with no email address. 2. **Add prepaid credit.** The minimum top-up is $3. 3. **Create an application key.** Application requests use Bearer API keys. Account and key management stay in your signed-in web session, and inference keys can't reach account-management actions. 4. **Change the base URL and model.** Keep your OpenAI SDK, set the base URL to UncensoredLabs and use an approved model ID. The homepage's Python example is two lines of config: ```python base_url = "https://uncensoredlabs.dev/v1" api_key = os.environ["UNCENSOREDLABS_API_KEY"] ``` The same request works from cURL against `/v1/chat/completions` with `"stream": true`, and responses arrive as server-sent events. The site lists OpenAI, JavaScript, Python and Go SDKs, cURL and plain HTTP. For coding clients, the [coding agents guide](https://uncensoredlabs.dev/docs/coding-agents) covers any client that accepts a custom OpenAI-compatible Chat Completions provider, with setup pages for Cline, Continue, OpenCode and DeepSeek Harness. ![UncensoredLabs pricing section with $3 to $100 top-ups, Developer, Growth and Scale packages, and per-model token rates](https://res.cloudinary.com/dhdy0tbpt/image/upload/c_crop,w_1440,h_1400,x_0,y_6340/v1794677302/ShipBoost/screenshots/screenshot-51186fff-0f39-48cc-ae37-a86e00eb57ad.webp) ## Pricing and billing There's no subscription. You buy credit, each request is charged at the published input, cached-input and output rates, and the balance never expires. Prices are snapshotted per request: the charge is reserved before the call and settles on actual usage, and uncertain usage stays reserved until it's reconciled. One-time top-ups start at $3. Three optional packages add a temporary usage discount: | Package | Credit | Discount on token usage | Support | |---|---|---|---| | Developer | $20 | 2.5% for 30 days | Basic | | Growth | $50 | 5% for 30 days | Basic | | Scale | $200 | 10% for 30 days | High-priority | Each package is a single payment with no automatic renewal, and overlapping discounts don't stack. All packages cover every model in both Chat and the API. Payment is by crypto via OxaPay, or by manual payment from $25 with a 10% fee. ## How UncensoredLabs handles your data UncensoredLabs says its application servers don't store prompts, model responses or tool-call content, and that prompts and responses aren't used to train models. What it does retain is usage, billing and security metadata, under its published policies. The [data controls page](https://uncensoredlabs.dev/docs/data-controls) sets out the split. Two details matter if retention is your reason for choosing a provider. First, the inference provider's own retention policy applies, and the model cards show this as "Provider policy applies". Second, browser chat history is saved only in your browser, and only if you opt in. Support tickets are stored separately: the subject and messages you submit, plus account and ticket identifiers, status and timestamps. ## Who it's not for - **Apps built on other API shapes.** The API supports OpenAI Chat Completions with text, streaming and tool calling. The OpenAI Responses API, Anthropic Messages, audio and embeddings aren't supported. - **Anyone who can't pay in crypto.** Top-ups go through OxaPay in crypto. The only other option listed is manual payment from $25, with a 10% fee. - **Anyone who needs a particular model.** The catalog is nine models at the time of writing, and new ones only appear after the operator verifies them. - **Anyone who needs zero retention end to end.** UncensoredLabs doesn't store your content on its application servers, but upstream inference-provider policies still apply. Use of the service is also governed by UncensoredLabs' terms and acceptable use policy, so read those before you build on it. ## FAQ **What does abliterated mean?** UncensoredLabs' catalog describes its abliterated models as having refusals removed from the weights. They're listed separately from the unmodified standard models. **What's the difference between uncensored and abliterated models?** UncensoredLabs lists them as two separate classes, but only its abliterated entries carry a description ("refusals removed from the weights"). Compare the individual model pages before you choose. **How do I run abliterated models without hosting them?** Call them through a hosted API. With UncensoredLabs, you add credit, create an application key, set your OpenAI client's base URL to `https://uncensoredlabs.dev/v1` and pick an abliterated model ID such as `glm-5.3-abliterated`. **Can I use abliterated models in a coding agent?** Yes, if the agent accepts a custom OpenAI-compatible provider. UncensoredLabs has setup guides for Cline, Continue, OpenCode and DeepSeek Harness, and tool calling works on models whose catalog entry lists it. **Are my prompts stored?** UncensoredLabs says its application servers don't store prompts or responses and don't use them for training. Usage, billing and security metadata are retained, and the inference provider's retention policy applies. ## Try it [Open UncensoredLabs](https://uncensoredlabs.dev/) to add credit and get an API key, or start with the [quickstart](https://uncensoredlabs.dev/docs/quickstart). You can also view the [UncensoredLabs listing on ShipBoost](/tools/uncensoredlabs), or browse more [API tools](/tags/api) and [AI platforms](/tags/ai-platform). --- Source: https://shipboost.io/blog/abliterated-models-openai-compatible-api-uncensoredlabs Markdown: https://shipboost.io/blog/abliterated-models-openai-compatible-api-uncensoredlabs.md