← sources

Questions Regarding Qwen 3.8 27B and Hallucinated Classes/APIs?

source: reddit (reddit:LocalLLaMA)published 2026-09-14 17:40ingested 2026-09-15 02:40

retained text

I've been running Qwen 3.8 27B at Q8 or BF16 on my CMP 170 - getting around 50 tps. Everyone raves about this model, but I am finding that it hallucinates a lot when asked questions that require some domain knowledge.

I am a developer that primarily works on a specific enterprise application that has its own Java classes, customization framework, etc. It's a common enough application that it would have a lot of training data from the web.

When I ask Deepseek (via the web interface) **or even Microsoft Copilot (not Github Copilot)** for assistance with specific code-related questions, I get code that works from them 99% of the time on the first try.

When I ask Qwen 3.8 27B, I get answers that contain hallucinated classes and APIs 90%+ of the time. It seems to really suffer when even a small amount of domain knowledge is needed.

**I assume this is because Qwen has been optimized for agentic use cases and not retaining domain knowledge in its weights - correct?**

Anyone had this issue?

If so, has anyone got any good tips for setting up a harness for Qwen - and perhaps a system prompt and/or specific skills - to discourage it from hallucinating made-up code and instead researching what it needs from the web before it responds? Or any other strategies?

**Update:** Will test more with opencode harness instead of llama.cpp web ui and see if that helps.

document fetched 2026-09-15 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1wgaapw/questions_regarding_qwen_38_27b_and_hallucinated