Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Alibaba's Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM, ships with a default reasoning effort of 'xhigh' that causes spectacular over-thinking. Simon Willison found it excellent but recommends raising the context limit to 262,144 tokens to avoid running out.
How this story unfolded
4 weeks · 17 reports · 173 community posts · 190 of 204 shown
- Aug 3
- Aug 11
- Aug 12
- Aug 13
- Aug 14
- Aug 15
- Aug 16
- Aug 17
- Aug 18
- Aug 19
- Aug 20
- Aug 21
- Aug 22
- Aug 23
- Aug 24
- Aug 25
- Aug 26
- Aug 27
- Aug 28
- Aug 29
- Aug 30
Qwen by email
Get an email when Qwen has news
No news that day, no email.
More stories today
- Redditor tests AI agents with $1 online task
- AI agents need their own identity before a gateway
- Claude Max users find default $200K spend limit
- TTFT-First Benchmark Ranks Lowest-Latency Voice and Realtime Agent APIs
- AI training demand causes Mac Mini shortages