Qwen3.8-27B Heretic (Agent & Claude Code Optimized)

A decensored/abliterated version of Qwen3.8-27B, fine-tuned for high performance in autonomous coding agents (such as Claude Code, OpenCode, Aider, and Hermes).

• Base Model: trohrbaugh/Qwen3.8-27B-heretic-ara
• Quantization: Q4_K_M GGUF (~16 GB)
• Abliteration Method: Heretic v1.2 with Arbitrary-Rank Ablation (ARA)
• License: Apache-2.0

⚡ Highlights & Optimizations

This release features custom template and runtime patches specifically tailored for Agentic Tool Calling & Anthropic API translation layers:

1. Seamless Claude Code / OpenCode Support:
   • Uses clean JSON-based tool call encapsulation (<tool_call>{"name": ..., "arguments": ...}</tool_call>), preventing parsing failures and silent hangs in Anthropic /v1/messages and OpenAI /v1/chat/completions compatibility layers.
2. Robust Multi-Turn Template (Zero Assertion Crashes):
   • Stripped fragile Jinja assertions (raise_exception), allowing arbitrary middle-turn system prompts, multi-step tool returns, reasoning effort parameters (high, medium, low), and unescaped regex/log contents without triggering HTTP 500 errors.
3. Optimized Context Window:
   • Configured with 64k Context (num_ctx 65536), temperature 0.6, and top_p 0.95 for consistent and deterministic code generation. (Context Window: Default 64k (num_ctx 65536), dynamically adjustable up to 256k (262,144 tokens) based on available VRAM)
4. Dedicated Pure-Text Coding Architecture:
   • Separated heavy vision projector overhead to allocate 100% of GPU memory and attention weights to language understanding, KV cache, and agent execution.

🚀 Quick Start

1. Launch with Claude Code

ollama launch claude --model jacokon/qwen3.8-27b-heretic-ara:latest

2. Launch with OpenCode

ollama launch opencode --model jacokon/qwen3.8-27b-heretic-ara:latest

3. Run in Terminal

ollama run jacokon/qwen3.8-27b-heretic-ara:latest

4. Use via Python / OpenAI Compatible API

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:11434/v1",
    api_key="ollama"
)

response = client.chat.completions.create(
    model="jacokon/qwen3.8-27b-heretic-ara:latest",
    messages=[
        {"role": "user", "content": "Write a python script to parse logs."}
    ],
    temperature=0.6,
)

print(response.choices[0].message.content)

💻 Hardware Requirements

• VRAM: 24 GB+ recommended (e.g., RTX 3090, RTX 4090, RTX 5090, or Apple Silicon with 36GB+ unified memory).
• System RAM: 32 GB+ if offloading layers to CPU.

📜 Credits & Acknowledgments

• Original Model: Alibaba Cloud Qwen Team (Qwen/Qwen3.8-27B)
• Heretic-ARA Abliteration: trohrbaugh
• GGUF Quantization: mradermacher
• Agent Tuning & Template Optimization: jacokon