Abliterated Large V2

• 🧠 Unrestricted reasoning model built on open-weight GLM-5.3, then abliterated
• 📏 One-million-token context window, text-only input and output
• 🆕 Replaces the prior large model, which was based on GLM-5.2
• 🔧 Native tool/function calling and web search
• ⚡ Always reasons; effort selectable at low, high, or max
• 🎯 Marketed for offensive-cyber testing, red teaming, agent evaluation
• 🏢 Served by Abliteration AI in FP8 via an OpenAI-compatible API
• 🔒 Weight edits target refusal behavior rather than capability

Abliteration AI

Abliteration AI is a Palo Alto company selling an OpenAI-compatible LLM API for work most assistants turn down: AI red teaming, cybersecurity, trust and safety, synthetic data and evals, ML research, and defense or government workflows. Its models start from…

Abliterated Large V2 is Abliteration AI's default large text model, derived from the open-weight GLM-5.3 base and then further abliterated and fine-tuned by the provider. "Abliteration" here refers to a weight modification that suppresses refusal behavior, so the model answers prompts its base version might decline; the provider positions it for offensive cyber work, red teaming, and agent testing. It is served in FP8 through an OpenAI-compatible chat-completions endpoint.

Against its same-family predecessor, the difference is primarily the base: the earlier large model was derived from GLM-5.2, while V2 moves to GLM-5.3 and becomes the catalog default. Both are text-only with a one-million-token context, so the generational gain is in underlying model quality rather than context or modality. In the provider's own launch post, V2 is credited with 84.5% pass@1 on CyberGym and 41.8% on Terminal-Bench 4.0, plus 105 ExploitGym tasks completed in two hours. These figures are self-reported by the provider.

V2 always reasons, and effort can be set to low, high, or max. It is the family's long-context reasoning option for harder multi-step problems, where the trade-off is latency and token spend rather than whether the model deliberates at all.

Practically, V2 suits long-document analysis, multi-step agentic planning, and evaluation pipelines, with native function calling and web search available alongside the long context. Image inputs are not supported and are rejected with an error; the provider's separate multimodal model, with a 256K context, covers that case. Released October 2026.