Abliteration AI sells a GLM-5.3 with the refusals edited out

Abliteration AI is now selling access to a version of Z.ai's GLM-5.3 with the refusal behavior stripped out. The bootstrapped startup opened a hosted API around August 29-31, 2026, branding the product abliterated-model-large-v2 and pitching it at offensive cyber operations and red teaming — the practice of attacking your own systems before someone else does.

The base model is Z.ai's GLM-5.3, a mixture-of-experts system with 753 billion parameters. Abliteration AI doesn't retrain it. It edits the weights — the numerical values that encode what a model has learned — targeting the specific refusal behaviors that make a standard model decline certain tasks. Customers never touch those weights. The company keeps them private and serves the model through endpoints compatible with OpenAI and Anthropic tooling.

CEO and co-founder Devon Thomas frames the whole thing as a commercial wrapper around a technique that already exists. The sell is convenience: a ready-to-use endpoint with options for customized enterprise policies. The model runs in FP8, a compact numerical format that trades some precision for speed and cheaper hardware, and carries a 1 million-token context window. Access runs $3 per million input tokens and $5 per million output tokens.

The performance numbers are all self-reported by Abliteration AI — not an independent evaluator — and concentrate on security and terminal work. On the privacy side, the company advertises strict zero data retention and asks only minimal identity verification. It says it does keep safeguards against child sexual exploitation and self-harm content.

Abliteration AI launched in late 2025. As of September 2026 it reported no venture funding and was running entirely on customer revenue. Wired described the broader practice of undoing model safeguards as something close to blasphemy in parts of the AI world, and Thomas answers the criticism on security grounds: defenders need to understand how attackers would use AI, and a model that refuses to engage with offensive techniques is of limited use for that job.

The approach is already spreading. Since the launch, other providers have started reselling or hosting similar abliterated GLM-5.3 variants. Those two design choices also sit in tension — zero data retention appeals to security firms handling sensitive client work, but combined with minimal identity checks it leaves the provider with little visibility into who is running the model, or why.