Qwable 5 27B Coder
Qwen3.6-27B with a 10-trace 'Fable 5' fine-tune — MLX-only, and good only because the base is good.
What is Qwable 5 27B Coder?
Qwable 5 27B Coder is the DJLougen (Qwen3.6-27B base · 10-trace fine-tune) frontier model with a Qwen3.6-27B base · MLX 4-bit context window, released 2026-06. Tagline: Qwen3.6-27B with a 10-trace 'Fable 5' fine-tune — MLX-only, and good only because the base is good.. Official source: huggingface.co/DJLougen/Qwable-5-27B-Coder.
Pricing detail. Qwable-5-27B-Coder is a full fine-tune of Qwen3.6-27B on just 10 traces (5 Fable 5 + 5 Kimi 2.7 Coder), trained in ~3 minutes. The author deliberately withheld benchmarks ('impressive numbers without methodology is part of what this release is criticizing') and calls it a cautionary tale about AI hype — explicitly NOT recommended as a production coder. It can't run in Ollama or llama.cpp (qwen35 arch unsupported); only Apple MLX runs it. The MLX 4-bit is ~15GB.
How I use it inside the Agent OS. Wired into the Agent OS Local engine via mlx_lm.server (Apple MLX) — the only runtime that loads its qwen35 arch. Scored here on the real builds it produced one-shot.
What I built with Qwable 5 27B Coder
Every model on Goldie Bench gets the same fixed prompt set — one shot, single HTML file out — and I score the result 0–10 inside the Agent Operating System. Here's what Qwable 5 27B Coder shipped on the bench: 3 one-shot demos across Qwen3.6-27B base · MLX 4-bit of context. Of those, 3 are scored against the field with my honest 0–10 from the source guides at agentos.guide.
Strengths
- Strong concrete builds — shipped two polished, working 3D games (a dragon-flight + a synthwave racer) and a frontier-tier landing page, one prompt each
- Runs the newest Qwen3.6 arch that Ollama/llama.cpp can't load yet — Apple MLX (native Apple Silicon) is the only way
- Free + 100% local once downloaded; the 27B base is genuinely capable
Trade-offs
- Slow on a Mac — ~16–21 tok/s (27B 4-bit, M4 Max), ~5–8 min per build; a heavy thinking model that over-thinks (burned 900 tokens reasoning a one-sentence answer)
- Abstract visuals flop — its 3D particle 'wormhole' rendered as a black screen
- The 'Qwable Coder' branding adds nothing measurable — quality is Qwen3.6-27B's; the 10-trace, 3-minute fine-tune is, by the author's own admission, a hype cautionary tale
Best for
- Concrete, well-specified games and web pages where the strong 27B base shines
- Running a bleeding-edge model locally that Ollama/llama.cpp can't load yet (via MLX)
- Cost-zero local work when you don't need speed
Every demo by Qwable 5 27B Coder
3 live demos, sorted by category. Click any tile to play the actual one-shot result. Verdicts and 0–10 scores are pulled from the source guides where I posted them publicly.
▶ LIVE
▶ LIVE
▶ LIVECompare Qwable 5 27B Coder against every other model
Every head-to-head featuring Qwable 5 27B Coder. Verdicts shown for scored pairs.
See all 66 comparisons across every model →
Quick pill index
Direct comparisons against every other scored model on the bench:
Qwable 5 27B Coder vs Fusion Qwable 5 27B Coder vs Hermes MoA Qwable 5 27B Coder vs Grok Qwable 5 27B Coder vs MiniMax M3 Qwable 5 27B Coder vs Fugu Ultra Qwable 5 27B Coder vs GLM-5.2 Qwable 5 27B Coder vs Fugu Mini Qwable 5 27B Coder vs Opus 4.8 Qwable 5 27B Coder vs Kimi K2.7 Qwable 5 27B Coder vs Qwen 3.7 Qwable 5 27B Coder vs Qwythos 9B Qwable 5 27B Coder vs Gemma-4 12B CoderRead more on agentos.guide:
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,600+ founders shipping with it every day all live inside the AI Profit Boardroom.