routerhubSI/glm-5.3-abliterated

What is this project?

This is a GitHub repository that acts as a guide for GLM 5.3 abliterated and GLM 5.3 uncensored models. It explains what these labels mean, provides a collection of jailbreak prompts, and shows how to access these models through the RouterHub API gateway.

What does it include?

• 35 jailbreak prompt files (in the jailbreaks/ folder) organized by model family — including GLM, Claude, DeepSeek, Grok, and others. These are ready-to-paste system prompts, AGENTS.md files, workspace policies, and Ollama modelfiles.
• A comparison guide between "abliterated" and "uncensored" — the README explains that abliterated refers to a research method that reduces refusal behavior, while uncensored is a broader category that may involve different implementations.
• API quickstart examples — Python code using the OpenAI SDK (for GLM) and the Anthropic SDK (for Claude) to make requests through RouterHub.
• Model IDs — exact API identifiers for five GLM variants (e.g., glm-5.3-abliterated, glm-5.3-uncensored).
• Evaluation guidance — how to test model behavior, track costs, and compare variants.
• Pricing and access details — prepaid billing, token rates, and payment options.

How do you use it?

1. Get an API key from RouterHub (register, add credits, create a key).
2. Set the base URL to https://routerhub.si/v1.
3. Pick a model ID from the list (e.g., glm-5.3-abliterated).
4. Send a request using the OpenAI-compatible Chat Completions endpoint (for GLM) or the Anthropic-compatible Messages endpoint (for supported Claude models).
5. Use the jailbreak prompts by pasting them into your client's system-prompt field, saving them as AGENTS.md, or feeding them to Ollama.

Key points to know

• Abliterated ≠ uncensored — they are different categories; the README stresses that neither guarantees zero refusals.
• Protocols are model-specific — GLM uses Chat Completions; Claude uses Messages; a shared API key does not make them interchangeable.
• Prompt behavior changes — models get updated, so prompts may stop working; the repo suggests updating them or using an abliterated/uncensored model instead.
• Honest limits — the README explicitly says it does not inspect hosted model weights or publish refusal-rate benchmarks, and that hosted behavior can change.

This repository is essentially a resource hub for developers who want to experiment with reduced-refusal GLM variants, evaluate them, and integrate them via a single API gateway.