Abliterated Large V2

• 🧠 Unrestricted reasoning model based on GLM-5.3, further abliterated and fine-tuned.
• 📏 One-million-token context window for long documents and large evaluation runs.
• 💬 Text-only — no image or audio input.
• 🎯 Always reasons; reasoning depth selectable as low, high, or max.
• 🔧 Supports function calling for tool-using and agentic workflows.
• 🌐 Web search is available as a capability.
• 🆕 Version 2 of the Abliterated Large line, released in 2026.
• 📚 Positioned for harder reasoning and evaluation workloads.

Abliteration AI is an organization focused on uncensored language modeling, taking its name from "abliteration" — the post-training technique that identifies and removes the internal direction associated with a model's refusal behavior, leaving the underlying…

Abliterated Large V2 is Abliteration AI's large text model, described by the catalog listing as an unrestricted reasoning model built on GLM-5.3 and then further abliterated and fine-tuned. Abliteration refers to a weight-editing technique that targets the internal directions associated with refusal behavior, so the resulting model declines fewer requests than the base release it was derived from.

The model is text-only and does not accept image input, but it carries a very large one-million-token context window, which suits long-document analysis, extended multi-turn sessions, and evaluation workloads that need many examples or traces in a single prompt. Its listed capabilities are reasoning, function calling, and web search, so it can be wired into tool-using and agentic pipelines rather than used only for plain chat.

Reasoning is always on: unlike models where thinking can be toggled off, this one always produces a reasoning pass, with the depth configurable as low, high, or max. That gives a simple dial between latency and thoroughness — lower settings for routine generation, maximum depth for harder multi-step problems.

As the second version in the Abliterated Large family, it is the successor to the earlier large model in the same line, released in 2026. No independently verified benchmark results for this specific hosted build are available from a qualifying evaluator, so capability should be assessed against your own workloads. Users should also note that reduced-refusal models shift safety responsibility onto the deploying application.