Open source repo/product/project built with open weight models?
Posted by rm-rf-rm@reddit | LocalLLaMA | View on Reddit | 7 comments
There are so so many "I built this flappy bird, tower defense, etc." social media posts for every model release, including open weight ones.
But is there a legit project/repo out there that was built wholly with open weight models? It doesnt have to be anything insane - even something basic like a VS Code extension etc. will do
Parzival_3110@reddit
The tricky part is proving "wholly built with open weights" in a way that is meaningful. A finished repo does not tell you whether the model wrote 90% of it, fixed two bugs, or just generated the first draft.
If I were looking for legit examples, I would trust projects that publish the process artifacts: prompts or agent logs, commit history, test failures, review notes, and which model/quant was used. Otherwise it is mostly a vibes claim.
Open weight models seem strongest when the repo has tight boundaries: CLIs, browser extensions, small dashboards, data cleanup tools, config generators, etc. The scaffolding matters more than the app category. A boring project with a reproducible local loop and tests is more interesting than another flashy demo.
Honestly this would make a good benchmark: pick a small real extension or tool, require only local/open-weight models, publish the full build log, and judge the result by tests plus maintainability rather than whether the first screenshot looks cool.
rm-rf-rm@reddit (OP)
Yup you nailed everything im thinking
TheRaiff1982JH@reddit
if you wanna try a new one try https://github.com/Raiff1982/Codette-Reasoning you might be surprised
zmarcoz2@reddit
I built this project (my favorite tts model) with deepseek v4 pro https://github.com/Cirius0310/echo-tts-cpp
SM8085@reddit
Define legit?
I like my stash extensions. whisper_transcribe uses a whisper.cpp backend to transcribe videos so you don't have to unmute them. llm_image_tag uses vision multimodal models along with the stash tags to try to figure out what tags to use.
sdfgeoff@reddit
Qwen 35B and 27B were the only models used for making this application:
https://github.com/sdfgeoff/server-monitor
Enough-Astronaut9278@reddit
Honestly I think most real projects using open weight models just don't market themselves that way. It's like asking "what projects are built with PostgreSQL." Nobody leads with that, it's just infra.
Also a lot of the serious usage is internal tooling at companies that will never be open sourced. The incentive to post about it is basically zero compared to "look I vibed a game in 5 minutes."