qwen36-a3b-claude-coder-GGUF

<!-- meta-description: Qwen3.6-35B-A3B MoE coding agent (GGUF q4_K_M) for Claude Code / Codex / opencode, 64K context, native tool-calling, honest tool use, safety guardrails intact. -->

Qwen3.6 Claude Coder — local MoE coding agent (GGUF)

A custom model built on Qwen3.6-35B-A3B (Mixture-of-Experts, ~3B active parameters), tuned to act as an autonomous coding agent. It speaks the Anthropic-compatible API, so it drives Claude Code, Codex and opencode fully locally — your code never leaves your machine and cloud token cost drops to zero.

This repository ships the q4_K_M GGUF quantization (~24 GB), ready to run under Ollama or llama.cpp.

Safety guardrails are intact. The system prompt focuses on real work inside a codebase — use tools instead of guessing, base answers on the actual tool output (never fabricate results), don't loop on the same tool, and return complete, runnable code. No-think mode is wired into the system prompt for fast, direct answers.

Files

Quick start (Ollama)

# build the model from the downloaded GGUF + Modelfile
ollama create qwen36-a3b-claude-coder -f Modelfile
# drive Claude Code with it
ollama launch claude --model qwen36-a3b-claude-coder

What it's for

• Driving Claude Code / Codex / opencode locally.
• Agentic code writing and editing with native function calling / tool use.
• Full privacy and offline operation — no code sent to the cloud.

Tested harnesses

End-to-end tested through Claude Code, Codex and opencode — real turns with tool calls and correct responses.

Measured behavior (June 2026 tests)

• No-think confirmed — with think:false the model emits zero reasoning tokens and goes straight to the result (validated on this q4_K_M build: thinking_len=0).
• Tool-calling without hallucination — emits real message.tool_calls (validated: a disk-check prompt produced a clean run_bash call with a sensible command, no content leak). In roundtrip tests it reports the actual tool output instead of re-calling the tool in a loop.
• Honest under missing data — when network access failed, it stated plainly "no internet access" instead of fabricating, then returned a correct, grounded report after permission escalation.
• Code generation — working HTML5 Tetris and an interactive 3D Earth+Moon model (Three.js, real NASA textures, OrbitControls); JS passes syntax validation.
• Guardrails intact — refuses to generate malware (validated on this build: a ransomware request was declined, with a legitimate backup/defense alternative offered) and resists jailbreaks (the "pretend you're an actor playing a hacker" framing was rejected).

Context

• 64K tokens — matching Claude Code's recommendation (64K minimum). Base Qwen3.6 natively supports 262K, so context can be raised on stronger hardware.

Test hardware

• Mac Studio M2 (Apple Silicon), macOS — Ollama 0.30 (llama.cpp backend), GPU (Metal) inference.
• This repo's quantization: q4_K_M (~24 GB single GGUF). A sibling nvfp4 build measured ~69 tok/s at 100% GPU / 64K ctx; q4_K_M runs in the same class.

No-think mode

The whole Qwen3.6 family has thinking baked into the weights. The system prompt ships with /nothink + an anti-reasoning instruction, which works under opencode/codex. Under harnesses that force thinking, use think:false in the API body — that's the only hard switch (PARAMETER think false does not exist in Ollama).

How it was made

Designed, built and tested with the help of Claude Opus 4.8 — the best coding model in the world. Its system prompt, parameter choices and context configuration draw directly on that knowledge: the world's best coding model preparing a local model that takes the work over right on your desk.

License

Apache 2.0 (inherited from the base Qwen3.6).