I'm also excited about the research that could be enabled by having weight-level access and fine tuning access on frontier open source models. There's a lot of interesting behavior that just doesn't exist in 8B parameter models, not to mention with architectural and training differences.
Some more details on this:
After realizing Hugging Face would be messy to work with to train Kimi-k2-thinking, we decided to do it ourselves.
We started with PrimeRL and implemented Kimi in it, verifying it against the Moonshot API. The initial distributed training method, FSDP, is not ideal for memory bottlenecked MoEs, so we added support for Expert Parallel. This enabled faster training, but many optimizations remained. We discuss several in the post, and collectively, these efforts took us from training 125 tokens/s to 6,660 tokens/s on a single 8xH200 node! Per token, our codebase is cheaper than anything on the market, including training APIs like Tinker.
We plan to open source in the coming week or two, pending safety evals!
I agree full transparency on data adds several other challenges. Still, even releasing the software and infrastructure aspects would be a huge step from where we are now. Also, some recent work has shown pretraining filtering to be possible and beneficial which could help mitigate some concerns of sensitive data in the datasets.
I'd also add training checkpoints to the list for active transparency. I think the Olmo models do a decent job, but it would be cool to see it for bigger models and for ones that are closer to state-of-the-art in terms of both architecture and algorithms.
yeah, the costs are definitely a factor and prohibitive in completely replicating an open source model. Still, there's a lot of useful things that can be done cheaply, including fine tuning, interpretability work, and other deeper investigations into the model that can't happen without the infrastructure.