AI BriefingThe Scaffold Saved It
Key point
With Qwen3.6 35B and a routing scaffold, 9 out of 10 real Go tasks passed.
Details
A routed local pipeline running Qwen3.6 35B on an RTX 5090 32GB via Ollama, with the larger Qwen3-Coder-Next 80B attached as an auxiliary repair/editor, passed 9 out of 10 real Go tasks.
The comparison results are as follows.
- GPT-5.4 best-of baseline: 10/10
- Routed local process: 9/10
- Qwen3.6 + little-coder: 8/10
- Qwen3-Coder 30B + little-coder: 5/10
- Existing local Gandalf harness: 3/10
The evaluation used 10 real tasks drawn from the author's Go repo, using a cloned workspace rather than the live repository. The task scope included CLI changes, dependency enforcement, an embedded version file, clock abstraction, error taxonomy, SQLite primitives, migrations, and baseline schema work.
The key conclusion is that swapping the model alone had limited effect, and performance rose significantly when scaffold, repair loop, and routing policy were combined.
The hardware setup is as follows.
- RTX 5090 32GB: Qwen3.6 35B A3B Q4_K_M
- RTX Pro 6000 96GB: served as the larger repair/editor role
- Qwen3.6 used about 27GB VRAM, allowing it to run alongside an embedding service.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
KakaoBank Hosts KodeRunner 2026, Presenting Agentic AI Vision Centered on 'Reliability'
Step 5 Preview Released: 600B Parameter MoE Model Highlights Agentic Tasks and Financial Analysis PerformanceStanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster Inference