Published October 5, 2026
Abliterated GLM 5.3 went live on Secret Chat AI on October 3, 2026. It replaced the short-lived model named Venice. The underlying model stayed the same, but Secret Chat AI now calls Abliteration AI directly and keeps Venice as a backup route. The result is a full model family with fast answers, deeper reasoning, web search, research and image creation.
What GLM 5.3 Is
GLM 5.3 is an open-weight language model from Z.ai. Z.ai released it on August 14, 2026, with an emphasis on coding, long-running agent tasks and cyber work. The company says it kept the same base model as GLM 5.2 and gained ground through more post-training. Its published evaluations use contexts as long as 1 million tokens for some tasks, though the exact setup varies by benchmark.
The weights are available from Z.ai on Hugging Face. They use the custom GLM 5.3 License, which permits use, modification, distribution, fine-tuning and derivative works subject to its conditions. That makes “open-weight” the accurate term.
What Abliteration AI Changed
Secret Chat AI runs Abliteration AI’s derivative, called Abliterated Large V2 by its maker. Abliteration AI says it started with GLM 5.3, then abliterated and fine-tuned it. The hosted model is text-only and has a 1-million-token combined input and output window. It supports low, high and maximum reasoning effort on Abliteration AI’s API.
Abliteration targets a model’s refusal behavior. The original research found a direction in a model’s internal activations that was strongly tied to refusals. Suppressing that direction made the tested models less likely to refuse, while adding it could make them refuse harmless requests. The paper tested 13 open models with up to 72 billion parameters. It did not test GLM 5.3, so it explains the method rather than proving how this particular derivative behaves.
The change is not cost-free in theory. Abliteration AI’s own documentation says capability can shift after the edit, including instruction following, tool use, coding and multilingual work. That is why a familiar model name does not guarantee identical answers from the original and abliterated checkpoints.
How GLM Abliterated Runs Here
In Secret Chat AI, the family is named GLM Abliterated. Smart Agentic chooses the route for you. Fast Chat uses low reasoning. The main reasoning route uses high effort. Web Search adds Abliteration AI’s search, while Research combines high reasoning with a broader search. Search answers include linked sources.
The family can also create pictures through Seedream. Attached pictures are described before the text model receives them. Documents are placed into the conversation as text. PDFs up to 10 MB are read as extracted text, so a scanned PDF without a text layer will not work; attach images of its pages instead.
Our October 3 check produced 417 tokens in 3.7 seconds, about 110 tokens per second. That was one internal measurement, not a speed promise. Load, prompt length, reasoning and provider conditions can all change the wait.
GLM Abliterated is paid. It also takes one seat in Duet Lax, where it answers beside Grok and an uncensored referee compares the two responses. For a closer look at the family, visit Private GLM Abliterated. The broader model guide is at Uncensored LLMs.
What Uncensored Means Here
GLM Abliterated uses Secret Chat AI’s Uncensored moderation level. Our screen is lighter than the Standard level, which helps with dark fiction, satire, political argument, anatomy and other lawful work that ordinary chatbots may reject too quickly. The floor still applies: sexual content involving minors, terrorism and real threats of violence are blocked on every model. Abliteration AI also keeps its own rules, so a request can still be refused.
The Privacy and Reliability Limits
Every prompt is screened first by OpenAI’s moderation classifier, without your account details attached. If it passes, the prompt, document or extracted PDF text goes verbatim to Abliteration AI. We store no chat archive on our side, build no profile from your chats and never use prompts for training. Your history stays in your browser.
Abliteration AI declares that prompts and answers are processed in memory, not stored by default and not used for training. Its web search is a separate limit: the query goes to an unnamed third-party search provider, and Abliteration AI’s no-storage statement does not cover that step. Remove names and other identifying details before sending. Secret Chat AI removes you from your queries; it does not remove the data from your messages.
If Abliteration AI is unavailable, Secret Chat AI sends the same model request through Venice without web search. The privacy report names Venice when that happens. If both routes fail, the request is not charged.
Frequently Asked Questions
- When did Abliterated GLM 5.3 arrive on Secret Chat AI?
It went live on October 3, 2026. This article was published on October 5, 2026.
- Who made GLM 5.3?
Z.ai made the base open-weight model. Abliteration AI created the abliterated and fine-tuned derivative used here. Secret Chat AI is not affiliated with either company.
- What did it replace?
It replaced the chat entry named Venice. Existing Venice chats continue on GLM Abliterated, with Abliteration AI as the main provider and Venice as backup.
- Can it search the web?
Yes. Web Search and Research use Abliteration AI’s search and return linked sources. The search provider is unnamed, and its handling is outside Abliteration AI’s no-storage statement.
- Can it read pictures and PDFs?
Pictures are described for the text model. PDFs up to 10 MB are sent as extracted text. Scanned pages need to be attached as images.
- Is GLM Abliterated free?
No. It is a paid model family. The app shows the cost under each answer.
- Does uncensored mean every request is accepted?
No. It means Secret Chat AI applies its lighter screen. The safety floor remains, and the provider can still refuse a request under its own rules.