Post

NEW POST @birgitta410.bsky.social recently spent some time trying out running local LLMs for some programming tasks. In this memo she outlines the factors that influence how viable they are for the job. martinfowler.com/articles/exp...

martinfowler.comViability of local models for codingNotes from my Thoughtworks colleagues on AI-assisted software delivery

7/7/26, 12:40 pm4 reposts24 likes

Andrea Grassi

@andreasdl.bsky.social

7 July

@birgitta410.bsky.social Regarding the context size, how did you notice you needed to move from 32K to 64K? Was the LLM just failing, or outputting hallucinated code?

Birgitta B

@birgitta410.bsky.social

7 July

Good question - I think I never let it come that far to find out because the harness (OpenCode or Pi) shows me how full the window is, or even refuses to continue if it’s full, or even start a session in the case of OC, if I remember correctly

Andrea Grassi

@andreasdl.bsky.social

7 July

Ok, it would be interesting to see what happens with the code quality when there's context compaction with smaller context sizes, where the harness could make a difference.

Andrea Grassi

@andreasdl.bsky.social

7 July

Obviously if the context is not enough to fit the request this would fail instantly, but for longer running task it could be an interesting datapoint.

Birgitta B

@birgitta410.bsky.social

7 July

Yes, composition is done by the harnesses. And when the context is actually full the API would return an error code that the harness has to deal with