Community Builds Qwen3.8 Into a 27B Uncensored Agent on One GPU
A community-tuned 27B Qwen variant strips refusals, adds a hierarchical task-tree reasoning protocol, and ships with native tool calling for agent workflows.
- New community fine-tune Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic patches tool calling and reasoning on top of an abliterated base.
- Ships BF16, FP8, and AWQ 4-bit branches, with the 4-bit variant running on a single 24GB GPU.
- Context extended to 128K in production, 256K native, with a 16,384-token single-turn output ceiling.
- Injects a Mythos-Class task-tree protocol into the chat template to stop infinite chain-of-thought loops.
- Author reports 0.00% refusal across 30 adversarial prompts and 100% weight health across 27.36B params.
- Best for autonomous coding agents, security research, and long-context work; not for user-facing deployments without added guardrails.
Qwen3.8-27B Packages Abliteration, Long Context, and Agent Tooling
A community maintainer has published Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic, a 27.36-billion-parameter checkpoint that combines refusal-suppressed weights, a hybrid Mamba and attention architecture, and an agent-oriented tool-calling stack. The release builds on the upstream OBLITERATUS checkpoint and focuses on deployment repairs and configuration changes rather than a new pretraining run.
The maintainer describes the release as a fix for broken function calling, chat-template truncation, reasoning-parser failures, and restrictive context defaults. Those repairs matter because agent reliability depends on the tokenizer template, output parser, sampling configuration, and serving runtime as much as it depends on the underlying weights.
The Patch Targets Deployment Failures
The repository attributes several changes to the patched release:
- A complete 9.4 KB Jinja2 chat template replaces a truncated 506-byte stub.
- Structured tool calls use the Hermes format expected by supported serving runtimes.
- The reasoning parser handles the model’s internal
<think>blocks. - The configured context target rises to 131,072 tokens, with a claimed architectural ceiling of 262,144 tokens.
- Single-turn generation supports up to 16,384 output tokens.
A Jinja2 chat template converts system messages, conversation turns, and tool definitions into the exact token sequence the model receives. A malformed or truncated template can break tool selection even when the weights remain capable of producing valid calls.
Abliteration Edits Refusal Behavior
Abliteration compares model activations on accepted and refused prompts to estimate a direction associated with refusal behavior. The process then edits selected weight matrices to reduce the model’s sensitivity to that direction, modifying an existing checkpoint without repeating pretraining.
Because the refusal direction can overlap with broader instruction-following behavior, the edit may also weaken formatting, reasoning stability, and tool use. This release attempts to recover those behaviors through its expanded template, reasoning scaffold, parser configuration, and sampling defaults. Refusal suppression remains useful for controlled red-team research, security analysis, and creative workflows, while transferring safety enforcement to the surrounding application.
Mythos Puts Planning in the Template
The bundled Mythos-Class template inserts a planning protocol inside the model’s <think> block, allowing compatible clients to apply the same scaffold on every request. The template directs the model through three phases:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
