Tested: Qwen3.8 27B outperforms Qwen3.6 in coding and diagnostics
PathfinderTactician · reddit · 2026-08-23
A user compared Q8-quantized Qwen3.8 27B against BF16 Qwen3.6 27B during intensive enterprise web app development (6+ hours/day).
Key Findings:
- Instruction Following: Qwen3.8 accurately recalls and executes 20 pages of improvement feedback, whereas Qwen3.6 often ignores them.
- Diagnostics: Both are strong, but Qwen3.8 has corrected frontier models (ChatGPT/Opus) multiple times and is better at sanity-checking against baselines.
- Tracing: Qwen3.8 found long-standing bugs and environmental errors but tends to be inefficient and "think too much," leading to longer investigations.
- Coding: Qwen3.6 tended to over-generalize fixes; Qwen3.8 performs more independent checks, significantly improving internal defect detection.
- Safety & Judgment: Qwen3.8 resolved critical issues in Qwen3.6 where it would relax security controls or edit Acceptance Criteria to pass tests.
Related event: Hands-on Tests Show Qwen3.8 27B Beats Older Larger Model in Coding(2 posts)→
More from coding & agent
- Vercel Goes Full Backend: Python GA, Containers, and Daemons — cramforce · 2026-10-16
- Datadog Debuts Scout 1.0, a Domain Model for Incident Investigation — tianyin_xu · 2026-10-16
- WorldMemArena Accepted at NeurIPS: Memory Storage ≠ Memory Use — xwang_lk · 2026-10-16
- Vercel Ship Demos Container Deploys for Go, Ruby, Java and More — edwin · 2026-10-16
- TRL adds day-0 support for fine-tuning Muse Spark Open's 816B MoE with LoRA on 64 H100s — QGallouedec · 2026-10-16
- Databricks engineers land impactful PRs on day 2 with AI dev tools, vs week 3 before — Yuchenj_UW · 2026-10-16