The post argues that Qwen 3.8 27B is one of the most impressive local models yet, not because it is perfect, but because a roughly 17 to 18 GB model can now do work that recently required a much larger hosted model. The catch is its default behavior. On xhigh reasoning, it burns huge numbers of tokens, overcomplicates simple requests, and can turn straightforward coding or generation tasks into long, expensive slogs. People who tried it broadly confirmed the same pattern. The model is genuinely strong, often close enough to much more expensive frontier systems to be surprising, but its shipped defaults make it look worse than it is.
The clearest consensus was that the problem is less "Qwen is bad" and more "the default harness is bad". Several people reported that low reasoning, medium reasoning, or even disabling visible reasoning entirely made the model much more usable while preserving most of the quality for ordinary tasks. Others noted that the default
GGUF template sets reasoning to xhigh, some tools do not expose the control cleanly, and runtime quirks like template handling,
KV cache behavior,
MTP support, and context settings can radically change performance. That turned model evaluation into a systems problem. The same weights can feel brilliant or unusable depending on the serving stack.
A second strong theme was just how far local inference has moved. Many readers were struck less by the overthinking bug than by the fact that laptops, Macs, and prosumer GPUs can now run a model this capable at all. People compared Qwen 3.8 to last year's top-end reasoning models and found the gap surprisingly small.
Apple Silicon came up repeatedly as a strong platform because large unified memory makes these models fit, even if bandwidth still hurts speed. Others argued that bandwidth, not raw compute, is now the real bottleneck for local inference, which is why
mixture-of-experts models often feel much snappier than dense ones of similar headline size.
The comments also landed on a broader point about current reasoning models. Overthinking is not an isolated Qwen quirk. Many see it as the natural result of reinforcement learning and benchmark incentives that reward exhaustive checking, long traces, and "proof of effort". That is great for coding benchmarks and agent evaluations, but lousy for day to day interactive use where speed and restraint matter more than squeezing out the last few benchmark points. People want models and tools that choose the right amount of reasoning automatically instead of defaulting to maximum effort. Until that exists, the practical advice was simple: tune the reasoning level, use templates that do not sabotage the runtime, and benchmark on your own workloads rather than trusting headline scores.
simonwillison.net