2026 Local Models: Analysis Over Generation

Local models such as Qwen 3.6 and Gemma 4 are close to being useful for Kotlin Multiplatform code generation but still fall short for many developers, according to Mark Murphy's experiments. Murphy suggests using local models for project analysis instead of code generation, citing the example of Qwen 3.6 via OpenCode identifying a dozen undocumented capabilities in kPointer in seven minutes. This approach offers incremental cost advantages, as local models only incur electricity costs after initial hardware investment.

Pretty much by definition, local models are less capable than frontier models. thank you for coming to my TED Talk Right now, in my experiments, the leading-edge local models Qwen 3.6 and Gemma 4 are close to being useful for Kotlin Multiplatform code generation. They might be perfectly fine for other stuff, but my guess is that the sheer breadth of software development options means that local models are still a bit short of the mark for many developers. However, code generation is not the only use of a coding agent harness and a model. One big advantage of a local model is incremental cost. You may have to pay a lot for the equipment to run the local model, such as my 64GB M2 Ultra Mac Studio. But, after that, the operating cost is mostly electricity. You are no longer paying per million tokens, whether directly or hidden behind a subscription model. So, one way to think about local models is: what can I do with cheap context that does not involve code generation? One angle is to give your local model agent harness access to suitable skills and have the local model analyze your project and report findings, rather than generate code. Not only does this give you "fresh eyes" virtually speaking on the project, but you can have the local model focus on aspects that you might not regularly spend the frontier model token on, such as: For example, while writing this newsletter issue, I asked Qwen 3.6 via OpenCode to take a peek at kPointer https://kpointer.commonsware.com : Analyze this project and its documentation in site/ . Are there significant capabilities in kPointer's implementation that are not covered in the documentation? Seven minutes and one second later, Qwen wrote up a dozen things that it thinks are missing or are under-documented. While a few suggestions are not that helpful, most seem legit, and I'll work to fold them in. That's with no specific documentation-review skills installed or anything other than a fairly generic OpenCode setup. You certainly can get the same sorts of output out of a frontier model, and possibly better. But, if you're aiming to minimize your use of frontier models, for financial or ethical reasons, the fact that you can get this sort of help from an open-weights local model is important. And it is likely that things will only get better from here, as open-weights models continue to improve. Have you had success with local models performing other sorts of analysis? Post a comment