GLM Abliterated is an uncensored GLM model for direct answers, long documents, web research and creative work. It is Abliteration AI's abliterated and fine-tuned version of the open GLM model from Zhipu AI, which publishes the family under the Z.ai name. Secret Chat AI reaches it through Abliteration AI's own API and puts its chat, search, picture and work-generator modes in one model family.
Z.ai publishes the underlying GLM family as open weights. Abliteration AI lists the model with a context window of up to one million tokens. In Secret Chat AI, attachments are capped per message and a very long chat is offered a summary so it can continue. A large context is not a promise that every detail will be recalled or interpreted correctly.
GLM is Zhipu AI's model family, published under the Z.ai name. Secret Chat AI is not affiliated with Abliteration AI, Zhipu AI or Venice.
What Abliterated GLM Means
In this setting, abliterated GLM does not mean a new model architecture. Research on open chat models found a direction in their internal activations that was closely tied to refusal behavior. Removing that direction reduced refusals, while adding it could make a model refuse harmless instructions. The researchers also warned that the change could produce unsafe answers and found a drop on one truthfulness test. “Abliterated” is therefore a description of changed refusal behavior, not a claim that the model is more accurate.
Abliteration AI also fine-tunes the GLM base it serves. Secret Chat AI does not perform that model work; it provides the access layer. For a fuller explanation of uncensored language models and their tradeoffs, read the uncensored LLM guide.
GLM Abliterated Modes in Secret Chat AI
You can leave the model on Smart Agentic and let a router choose how much work the question needs. If you prefer to decide, the mode button offers four clear paths:
- Fast Chat uses low reasoning for a quick answer.
- Web Search uses Abliteration AI's search and returns linked sources.
- Research spends more time reasoning and searches more widely.
- Create Image turns your description into a picture with Seedream.
The thinking seat uses high reasoning when a question needs it. Search has one important privacy exception: Abliteration AI passes the search query to an unnamed third-party search provider. Its no-storage statement does not cover that step. If that matters, use Fast Chat and leave out details the search provider should not receive.
Pictures, Documents and Scanned PDFs
GLM Abliterated can take pictures and documents alongside your message. Pictures are described for the model. Documents are inlined as text. That makes it useful when the question depends on a chart, a draft or several pages of source material rather than a short prompt.
PDFs take a specific route: Secret Chat AI extracts their text and sends that text to the model. A scanned PDF with no text layer cannot be read this way. Attach pictures of the relevant pages instead. The picture description may miss fine details, so check the source before relying on an answer.
When you ask this model to draw, Seedream makes the picture. GLM Abliterated handles the request and prompt, but it is not itself the image model.
An Uncensored GLM With a Safety Floor
GLM Abliterated appears on the Uncensored level and carries a flame in the model selector. Its screen is lighter than the Standard level, which gives fiction, satire, dark themes, clinical questions, political analysis and other lawful work more room. It still has firm limits.
Anything sexual involving minors, terrorism and real threats of violence are blocked on every model. Abliteration AI also keeps its own rules, so a request may still be refused. For the wider model list and the distinction between the two levels, see the uncensored AI hub.
Use It in Duet Lax and Work Generators
GLM Abliterated is one half of Duet Lax, the uncensored AI Council entry marked with a flame. It answers the same question as Grok. An uncensored referee then compares the two answers. Agreement can be useful evidence, but it is not proof; two models can make the same mistake. Read more about the comparison format on the AI Council page.
The family can also run all 11 work generators: Presentation, Chart, Table, Document, Resume, Email, Translator, Rewriter, Article, Social Post and Logo. In Smart Agentic chat it can build a deck, table, chart, document, post or article and hand the result to the matching editor. Its presentation and social-post pictures come from Seedream. Browse the work generators for the available outputs.
How the Abliterated Model Falls Back to Venice
The primary route goes straight to Abliteration AI. If that API fails because of an outage, rate limit, authorization problem or timeout, Secret Chat AI can send the same model request through Venice instead. The Venice fallback does not use web search, and the privacy report names Venice when it handled the answer. If both routes fail, you are not charged.
Privacy, With the Limit Stated First
Every prompt, picture and document goes verbatim to third parties. The OpenAI moderation classifier sees the request before the model provider does. Abliteration AI then receives it, and Venice may receive it on a fallback. Search adds the unnamed search provider. Remove names and sensitive details before sending; that is your responsibility.
Secret Chat AI removes you from your queries — it does not remove the data from your messages. We keep no chat archive on our side. Your history stays in your browser, where it is not encrypted at rest and does not sync between devices. We build no profile of you, and providers see our service rather than the person asking. Your requests are never used for training.
Abliteration AI says prompts and outputs are not retained by default, while operational telemetry such as token counts, timestamps and error codes may be kept. Its public page also says prompt content is not used as a training signal. Those are provider declarations, not a hardware audit. Each answer includes a Session Privacy Report that names the provider and its stated terms.
Price and Access
GLM Abliterated is paid. Usage comes from a prepaid dollar balance, with packs from $12. There is no subscription, and the balance never expires. The app shows the answering tier and exact charge beneath each answer. A failed provider request costs nothing.
See the full private AI model lineup if you want to compare GLM Abliterated with the Standard and other uncensored families.