Comment by jart
18 days ago
Gemma 27b can write working code in dozens of programming languages. It can even translate between languages. It's obviously not as good as Gemini, which is the best LLM in the world, but Gemma is built from the same technology that powers Gemini and Gemma is impressively good for something that's only running locally on your CPU or GPU. It's a great choice for airgapped environments. Especially if you use old OSes like RHEL5.
It may be sufficient for generating serialized data and for some level of autocomplete but not for any serious agentic coding where you won't end up wasting time. Maybe some junior level programmers may find it still fascinating but senior level programmers end up fighting with bad design choices, poor algorithms and other verbose garbage most of the time. This happens even with the best models.
> senior level programmers end up fighting with bad design choices, poor algorithms and other verbose garbage most of the time. This happens even with the best models.
Even senior programmers can misuse tools, happens to all of us. LLMs sucks at software design, choosing algorithms and are extremely crap unless you exactly tell them what to do and what not to do. I leave the designing to myself, and just use OpenAI and local models for implementation, and with proper system prompting you can get OK code.
But you need to build up a base-prompt you can reuse, by basically describing what is good code for you, as it differs quite a bit from person to person. This is what I've been using as a base for agent use: https://gist.github.com/victorb/1fe62fe7b80a64fc5b446f82d313..., but need adjustments depending on the specific use case
Although I've tried to steer Google's models in a similar way, most of them are still overly verbose and edit-happy, not sure if it's some Google practice that leaked through or something. Other models are way easier to stop from outputting so much superfluous code, and overall following system prompts.
I've spent a long time with models, gemma-3-27b feels distilled from Gemini 1.5. I think the useful coding abilities really started to emerge with 2.5.
The technology that powers Gemini created duds until Gemini 2.5 Pro; 2.5 Pro is the prize.