They claim several times to be “truly open” but I don’t see anything about open sourcing the training code. The inference code isn’t that interesting. What we need is total transparency on how the model weights are produced, since otherwise it’s hard to trust the biases of a model. The only actually truly open model is AI2’s OLMo as far as I know - and even they don’t get totally transparent about how they produced their training data set, which includes curation and filtering by “safety and ethics” people:

https://blog.allenai.org/hello-olmo-a-truly-open-llm-43f7e73...

But until training data sets and source code are released under an OSI license, Snowflake should stop with the open washing.

It might not be obvious from the title, but this model is absolutely massive: 480B parameters. The largest open-source model to date, I believe.

You can try it out here: https://arctic.streamlit.app/

Weights are here: https://huggingface.co/Snowflake/snowflake-arctic-instruct

I can still change the title, but I'm not sure if that would be okay according to HN guidelines.

Indeed – I was doing the editorialization in the comments. ;)

It's missing the name though (Arctic). That might be worth adding.

Done!