Liquid Foundation Models - 3 LLMs (1B, 3B, 40B) on Custom Architecture

[@timothywcrane]

Love new archs. Wish I could touch it. Benchmarks are generalized, the zero use case.

[@fahdmirza]

thanks for the feedback.

[@mdubbau]

Wish they were releasing as open source

[@fahdmirza]

yeah agreed

[@JeevaPrakash-t2h]

really appreciated.

[@fahdmirza]

Glad to hear that

[@simonjoseph551]

It is a good development but still needs to be worked on. It is trained to 2021. I asked the model how many c do we have in hippopotamus and it said 2.

[@fahdmirza]

Thats a really interesting test. I will use it in my tests in futures , thanks

[@gavin3405]

Awesome stuff! Can't count the r's in "strawberry" :(

[@fahdmirza]

yeah, true.

[@justtiredthings]

Qwen2.5 32b gets considerably better benchmark results, and it's a smaller model 🤷

[@fahdmirza]

thanks for feedback.