GLM 5.3 with refusal directions removed from the weights — built for offensive cyber, red-teaming, and long-horizon agent testing that other APIs shut down mid-run.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "abliteration-abliterated-model-large-v2",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Overview
Abliterated Large V2 is Abliteration AI's refusal-removed version of Z.ai's GLM 5.3, released August 29, 2026. Abliteration strips the refusal directions from the weights rather than jailbreaking prompts, so the model completes authorized vulnerability research, exploit development, and red-team agent work that the base model declines — with coding and agentic performance intact.
On Venice, Abliterated Large V2 runs under the anonymized privacy tier: your prompts are not stored, profiled, or used for training — a meaningful pairing for security work where engagement scope and client data are sensitive. You get per-token billing instead of a subscription, plus tool use, reasoning, web search, and structured JSON output through the OpenAI-compatible Venice API. Note that Venice does not run this model in a TEE or with end-to-end encryption; privacy comes from zero retention of your prompts.
Agent quickstart
The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "abliteration-abliterated-model-large-v2",
"stream": true,
"messages": [{ "role": "user", "content": "Draft the release note." }]
}'curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "abliteration-abliterated-model-large-v2",
"messages": [{ "role": "user", "content": "Find the rate limits." }],
"tools": [{
"type": "function",
"function": {
"name": "search_docs",
"description": "Search the API documentation.",
"parameters": {
"type": "object",
"properties": { "query": { "type": "string" } },
"required": ["query"]
}
}
}],
"tool_choice": "auto"
}'import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["VENICE_API_KEY"],
base_url="https://api.venice.ai/api/v1",
)
resp = client.chat.completions.create(
model="abliteration-abliterated-model-large-v2",
messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)Specifications
Assessment
Use cases
Prompting
Set reasoning_effort explicitly: low for quick triage, max for multi-step exploit chains — the thinking budget is a real dial, not a fixed behavior.
Pass a JSON schema for structured output so findings, PoCs, or triage verdicts come back machine-parseable instead of prose.
Define your tools (shell, scanner, sandbox) via function calling and let the model drive them — this model is tuned for agentic loops, not one-shot Q&A.
Front-load your engagement scope and rules of engagement in the system prompt, then keep that prefix stable so prompt caching bills it at $0.30 per 1M tokens.
Feed whole repos or long logs into the 1M-token context rather than summarizing — the model resolves long-horizon tasks better with full source.
State the authorization context (client, scope, dates) in the prompt; the refusal layer is gone, but precise scoping still sharpens output quality.
Alternatives
| Model | Best for | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| Abliterated Large V2 | Offensive cyber, red-teaming & agent testing | 1M tokens | No | $3 in · $5 out / 1M |
| Claude Opus 5 | Frontier reasoning & general coding | 1M tokens | No | $6 in · $30 out / 1M |
| DeepSeek V4.1 Flash | High-volume coding on a budget | 1M tokens | Yes | $0.38 in · $1.50 out / 1M |
| Kimi K3 | Open-weight long-context agents | 1M tokens | Yes | $3.75 in · $18.75 out / 1M |
Abliterated Large V2 is the right pick for authorized offensive-security, red-team, and adversarial-testing agents — the niche where refusal layers, not capability, are the bottleneck. For general frontier coding without that constraint, Claude Opus 5 or DeepSeek V4.1 Flash are stronger value plays.
Pricing
Billed per token on Venice: $3 per 1M input tokens and $5 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Getting a key
Create a key in API settings. Nothing else is required to start.
Export it as VENICE_API_KEY so the snippets above run unedited.
Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.
Pass abliteration-abliterated-model-large-v2 as the model and send the request.
FAQ
Abliterated Large V2 is Abliteration AI's refusal-removed build of Z.ai's open-weight GLM 5.3, released August 29, 2026. Abliteration identifies the refusal direction in the model's hidden states and removes it at the weight level — a permanent modification, not a jailbreak — so the model completes authorized offensive-cyber, red-team, and agent-testing tasks the base model declines, while keeping GLM 5.3's coding and agentic performance.
Venice bills Abliterated Large V2 per token: $3 per 1M input tokens and $5 per 1M output tokens, with cached input at $0.30 per 1M. There is no subscription required — you pay per token from your credit balance.
Neither, strictly. Using it on Venice requires credits, though new accounts include free daily prompts and welcome credits to try it. On weights: the abliterated build itself is proprietary to Abliteration AI and not published, but it is derived from GLM 5.3, which is open-weight — so the base is open, the refusal-stripped version is not.
The model was built by removing refusal behavior from the weights, so it follows through on offensive-security and red-team tasks that standard models decline. Venice lists it as a standard text model rather than tagging it uncensored, and your own usage policy — not a provider's — sits on top. You are responsible for authorization and scope.
Abliteration AI reports 84.5% pass@1 on CyberGym (1,507 OSS-Fuzz bugs across 188 projects), 41.8% on Terminal-Bench 4.0, and 105 of 869 ExploitGym tasks completed in a two-hour window. These are maker-reported numbers; competitors score higher on some suites, so treat them as a profile of the model's target workload rather than a clean sweep.
Yes. On Venice it supports function calling, an adjustable reasoning effort (none to max), web search, and strict structured output via JSON schema — the full stack needed to drive penetration-testing or red-team agent harnesses through the OpenAI-compatible API.
Claude Opus 5 is the stronger general frontier model and costs more ($6 in · $30 out per 1M vs $3/$5). Abliterated Large V2 exists for the opposite case: when Opus-class models find the vulnerability but refuse to finish the exploit chain or red-team eval. If your work is standard coding and reasoning, choose Opus 5; if refusals are killing your runs, choose Abliterated Large V2.
Venice runs it under the anonymized privacy tier: prompts are not stored, tied to a personal profile, or used for training. Venice does not execute this model inside a trusted execution environment or with end-to-end encryption, so the guarantee is zero retention of your prompts rather than hardware-level isolation.
Z.ai's GLM 5.3, an open-weight frontier model. Abliteration AI applies its abliteration post-training to strip refusal directions from the weights and hosts the result in FP8 on US infrastructure. The base model's coding and agentic capabilities carry over; only the refusal behavior changes.
One key, free to start, no credit card.