Qwen3.8-27B One Week Later: Community Verdict & Benchmarks
Jonathan_Rivera · reddit · 2026-08-23
Compiled from 2,000 posts, this is a deep dive into Qwen3.8-27B's performance one week post-release.
Key Takeaways:
- Consensus Pick: The new standard for local agentic coding, backed by tool-calling reliability rather than just benchmarks.
- Reasoning Presets: The default 'xhigh' thinks deeply but is slow. 'Low' and 'medium' presets score similarly on intelligence (43/44) while cutting thinking tokens by 7–9x and wall time by 6–7x.
- Trade-offs: Knowledge recall regressed vs. 3.6, likely due to agentic design choices.
- Quantization: Q4KM is indistinguishable from Q8 on perplexity, but real-world reasoning splits below Q6.
Strengths:
- Agentic Coding: Successful runs with 80 automated tool calls and 1M+ token contexts.
- Creative Coding: One-shot generation of playable game clones (Super Mario, Galaga) and an overnight MOBA build.
- Vision: Native support works well but is VRAM-heavy on 16GB cards.
More from coding & agent
- LORE-0: An autonomous agent foundry that finds capabilities or builds them — Inner_Oil706 · 2026-08-23
- Concept: Local dashboard assembled on-demand by your agent — irvinebroque · 2026-08-23
- An Upcoming Open-Source Tool for Testing AI Agent Behavior Before Production — GeologistRare8364 · 2026-08-23
- 11 Grok Bot tips: CEO agents, reverse prompting, and plugin workflows — AICopyLab · 2026-08-23
- Notion Aims to Build Enterprise-grade AI Skills Library for Better Agent Reuse — thisiskp_ · 2026-08-23
- Replit ships Free Mode, GitHub Skill import, and more updates — amasad · 2026-08-23