license: apache-2.0
tags:
- deprecated
- do-not-use
- cautionary-tale
⚠️ DEPRECATED — DO NOT USE THIS MODEL
This model does not work for real coding tasks. I benchmarked it in June 2026 and it failed essentially every test — one-shot and agentic. I'm leaving it up only as a documented cautionary example so nobody wastes a 42 GB download. Use a normal coder model instead (e.g. Qwen3-Coder-30B-A3B-Instruct).
What this was
An MLX 4-bit conversion of huihui-ai/Huihui-Qwen3-Coder-Next-Opus-4.6-Reasoning-Distilled-abliterated — a Qwen3-Coder-Next 80B (A3B MoE) with Claude Opus 4.6 reasoning distilled in, then abliterated. On paper it sounds great: "a coder that reasons like Opus." In practice it cannot stop reasoning long enough to actually finish anything.
The benchmark — it lost everything
Head-to-head against the plain Qwen3-Coder-30B on the same machine (Apple Silicon, MLX), same prompts:
| Task | Qwen3-Coder-30B | This model (80B reasoning) |
|---|---|---|
| Asteroids game (one-shot HTML) | ✅ working game | ❌ wrote a planning monologue, never produced the game |
| Snake game | ✅ 14s | ✅ but 8× slower, ~7× the tokens for the same result |
| Calculator | ✅ clean, complete | ❌ bloated to 56 KB and hit the token cap unfinished |
| Analog clock | ✅ working | ❌ bloated to 56 KB and hit the token cap unfinished |
Hard expression parser (18 hidden tests, no eval()) |
✅ 16/18 | ❌ 0/18 — produced no code at all, just an unclosed <think> block |
| Agentic task: write code, run it, fix until tests pass (with tools) | ✅ completed | ❌ called zero tools, wrote zero files, just said "DONE" |
Why it fails
A reasoning model distilled onto a coder is the worst of both worlds: it can't stop reasoning to converge. On one-shot tasks it over-generates and exhausts its token budget before finishing the answer. In agentic mode — its supposedly superior use case ("tool calling + bug detection") — it won't even pick up a tool; it narrates what it would do and declares itself done. The smaller, plain 30B coder beat it on speed, on completion, and even on the hard reasoning task.
Use instead
Any standard coder, e.g. Qwen3-Coder-30B-A3B-Instruct.
I converted this on reputation — "Opus reasoning distilled into a coder!" — before actually testing it. Lesson learned: benchmark before you adopt. Sharing the result so you don't repeat my mistake. — divinetribe
Part of Claude Code Local
This model is one of the fighters in Claude Code Local (3.2k★), which runs Claude Code 100% on-device on Apple Silicon through an MLX-native Anthropic-API server. Not sure which local model to run as an agent? Check the Agent-12 local agent leaderboard: real agent tasks, judged by the filesystem, same hardware for every row.
Built by Matt Macosko in Arcata, CA. Open to work on local-AI and Apple Silicon inference: [email protected].