Qwen 3.8 surprises from overnight testing

Article automatically generated from technical news.

I've been benchmarking Qwen 3.8 and it's competitors since last night, including Qwen 3.6. I got some unexpected results. Qwen 3.8's architecture seems to be identical to 3.6 and 3.5. It looks like Qwen 3.8 is primarily a training data change. Qwen's notes and other articles seem to support this, YMMV. Qwen 3.8's training data seems very narrowly tailored to a handful of scenarios. There was clearly a lot of expense and time put into benchmarking above everything else. I have duplicated t