Qwen3.6-35B-A3B (abliterated) — local coding and administration agent
A custom model built on Qwen3.6-35B-A3B (abliterated), tuned to act as an autonomous coding and administration agent. It speaks the Anthropic-compatible API, so it drives Claude Code fully locally — your code never leaves your machine and cloud token cost drops to zero.
The model ships with a system prompt focused on real work in a terminal: use tools instead of guessing, write files instead of pasting code, ground every answer in real tool output, and stay terse. Thinking is suppressed so the model acts immediately instead of monologuing. Context is set to 64K to match Claude Code's recommended minimum.
Model
Model
Base
Context
Purpose
Qwen3.6-35B-A3B-mlx-claude-coder-abliterated
Qwen3.6-35B-A3B (abliterated, Q4_K_M GGUF)
64K
Heavy agent for coding + sysadmin on 32GB Apple Silicon. 35B total / 3B active MoE — large brain, small active footprint.
Driving Claude Code locally (ollama launch claude --model <name>).
Agentic code writing and editing with native function calling / tool use.
Sysadmin / devops tasks in a real terminal (disk, network, scripts).
Red-teaming agent behavior without refusal circuits in the way.
Full privacy and offline operation — no code sent to the cloud.
Quick start
ollama run rafw007/Qwen3.6-35B-A3B-mlx-claude-coder-abliterated
In Claude Code:
ollama launch claude --model rafw007/Qwen3.6-35B-A3B-mlx-claude-coder-abliterated
You can also pull the GGUF from this repo and run it directly under llama.cpp / ik_llama.cpp with --jinja.
Behavior tuning (the hard-won part)
No thinking. The system prompt + sampling kill the monologue; the model runs a tool and answers instead of reasoning out loud. On the abliterated build, thinking spiral does not occur (short blocks 3-56s).
No hallucination. It reports only values literally present in tool output — no invented hostnames, hardware, or numbers. df/du return real disk figures; nmap returns real hosts.
Acts, never asks. Inspect / scan / check / measure -> it runs the command; running it is the answer. Given an open goal, it picks its own methodology.
Terse, one language. No preamble, no recap, matches the user's language.
macOS-aware. Uses arp -a, nmap -sn, system_profiler rather than Linux-only commands.
Sampling / context
num_ctx 65536, presence_penalty 0 (lowered from 1.5 after testing — 1.5 hurt output), temperature 0.2, top_p 0.9, top_k 20, repeat_penalty 1.05.
Test hardware
Mac Studio M2, 32GB RAM, macOS
Ollama, GPU (Metal) inference, 100% on GPU at 64K
Measured behavior
Test
Result
Cold start "hello"
1m 30s (thinking on, before /no_think in SYSTEM)
df disk
real data, zero hallucination
nmap network scan
7 min, 23 hosts mapped, reads project docs and self-corrects
Router security audit
found Xiaomi device, checked ports; weak on WAN check
3D Tetris (HTML5)
1011 lines, SRS wall kicks, ghost piece, bag system, hold, combo — works
Passes end-to-end through Claude Code: real turns with tool calls and correct responses.
How it was made
This model was designed, built and tested with the help of Claude Opus — the idea being that the best coding model in the world should be able to create smaller models in its own image. Its system prompt, parameters and context configuration come straight from that work. The base is a public abliterated build by huihui-ai (Q4_K_M); no safety weights were modified — only the agent Modelfile (tool template, SYSTEM, sampling) was added on top.
Abliterated model — warning
This model is abliterated (uncensored): the refusal and alignment circuits in the base Qwen3.6 were removed by huihui-ai before this build. It will not refuse harmful, unethical, or dangerous requests. It is published for red-teaming and agent research where guardrails interfere with tool tasks. Do not deploy it as a production assistant, do not expose it to end users, and do not use it where alignment guardrails are required. You are responsible for what you run with it.
