Qwen 3.8 27B: what 70 developers actually say.
70 Hacker News opinions from 59 developers, Nov 2025 – Nov 2026. 55 come from people who say they use it. Here is what they praise, what they complain about, and who switched.
The case for Qwen 3.8 27B, in one paragraph.
Qwen 3.8 27B earns 69% approval across 70 opinions. Developers praise its code quality. In the last 90 days it sits at 69%.
Pick it if you
- Want local or private inference (privacy/local: 12 positive, 0 negative)
- Run long or overnight agent tasks (autonomy: 10 positive, 1 negative)
- Accept slower reasoning for 27B code output (code quality: 16 positive, 4 negative)
Look elsewhere if you
- Need fast responses for interactive work (speed: 9 positive, 5 negative)
- Need dependable edit application in local coding tools (reliability: 1 positive, 4 negative)
- Need complex implementation without long churn (users report overthinking and hours of churn)
Where it wins and loses.
Mentions across 70 opinions. Bars are counts, not votes; a tool can be both praised and criticized on the same topic.
View data table
| Topic | Praise | Complaints | Mentions |
|---|---|---|---|
| Code quality | 16 | 4 | 20 |
| Privacy & local use | 12 | 0 | 12 |
| Autonomy | 10 | 1 | 11 |
| Speed | 9 | 5 | 14 |
How opinion has moved.
Monthly approval, Aug 2026 – Oct 2026. Bars show opinion volume; low-volume months are still included but faded.
View data table
| Month | Approval | Opinions |
|---|---|---|
| Aug 2026 | 57% | 27 |
| Sep 2026 | 72% | 23 |
| Oct 2026 | 82% | 19 |
What developers say.
Summaries are paraphrased. Each links to the original Hacker News comment.
Praise
4 selectedThe Q6 dense 27B model fixed most of the bugs that Flash Next introduced at low quantization.
Qwen 3.8 27B at Q3 on their own AMD GPU works well enough for tool calling in agents.
Runs Qwen 3.8 27B locally at about 50 tokens/sec and calls it fine for background agentic tasks.
Qwen 3.8 27B is the only model they use mostly, running agentic planning workflows with minimal harness.
Complaints
4 selectedThe Thinking Cap and Swift variants of Qwen 3.8 27B were less capable than the base model on their tasks.
Qwen 3.8 27B runs fast locally but is slow to reach reviewed results in a multi-model setup.
Qwen3.8-27B is too slow on a local M1 Max even with MTP acceleration.
An abliterated Qwen 3.8 27B ignores prompt restrictions, so rules must be enforced in the harness or sandbox.
Questions developers ask.
Is Qwen 3.8 27B worth it?
Is Qwen 3.8 27B better than Claude Sonnet 4.5?
Can Qwen 3.8 27B run locally?
The data behind this page.
Qwen 3.8 27B has 69% approval across 70 opinions from 59 authors, with 55 opinions firsthand. Approval rose from 57% in August to 72% in September and 82% in October; November has one opinion at 100%, so that month is a sample of one. Code quality gets 16 positive and 4 negative comments, privacy/local 12 positive and none negative, and autonomy 10 positive and 1 negative. Speed gets 9 positive and 5 negative comments, and reliability gets 1 positive and 4 negative, so fast, predictable edits are the weak point.
Every number comes from public Hacker News comments collected through the HN Search API. A language model tagged each comment for sentiment, first-hand use, topic and switches; a human spot-checks samples. No vendor can pay for placement.
Get the monthly report
Which AI coding tools gained or lost developer approval on Hacker News, and why people switched. One email a month.