This past week or so, I have been spending some time with the latest hotness in the world of local models: Qwen 3.8.
Some have described it as having capabilities on par with Claude Opus. I wouldn't go that far. However, in terms of what it can generate in an agentic coding scenario, the output is somewhere in the Claude Haiku-to-Sonnet range. This is a substantial improvement over the results from any local model that I have tried before, let alone previous editions of Qwen.
How it seems to get that quality is by thinking a lot.
If you use a coding agent harness that lets you view the thinking messages (e.g., OpenCode), you can see how many thinking messages it takes for Qwen 3.8 to get to the point where it takes an action. While the output token throughput seems reasonable for a local model, the sheer number of tokens that it takes means that while Qwen 3.8 delivers good results, it takes a minor eternity to get there. I have tried both the nvfp4
and mxfp8
encodings on my 64GB M2 Ultra Mac Studio, and Qwen takes 15-20 times as long to do some work as it would take Claude Sonnet. I had hoped that the smaller nvfp4
model might be faster, but it did not have an obvious effect.
Quoting Simon Willison's take: The fact that a 17GB file can do all of this stuff on my home machines is a miracle. Once again, I’m delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models—today it can run on a capable laptop.
The only thing holding this back from being a daily driver is performance. It feels pretty slow on both the M5 Mac and the DGX Spark. That’s the catch with these dense (non-Mixture-of-Experts) models—they require a whole lot of memory bandwidth to perform well, and neither of the machines I have access to are top performers in that regard.
The most important thing about Qwen 3.8 27B is
what it demonstrates. We can have an open weights general purpose model with a long context, effective tool calling, strong vision ability, and competent code generation, and we can fit the whole thing in just a 17GB file.
Qwen 3.8 still does well at analysis and does not exhibit the over-thinking behavior seen with its code generation. That suggests a few avenues for improving the pace of code generation:
Qwen 3.8 is not the "magic bullet" that I had hoped, but it demonstrates clear progress towards having an effective agentic coding experience using a local model. Considering how far we have come in the past 8 months, this is all very promising, and I keenly await future generations of this technology.