Rui Carmo
The timing for this is great, as I’m starting to get back to shoving LLMs into single-board computers. The 128K token context length seems to be becoming a standard of sorts (which is also nice), and I might actually try the vision models as well (with the usual caveats about SBC NPUs being quite limited in both TOPS and data formats).
And, of course, the “open” part is… Interesting. I’m becoming somewhat skeptical as to anyone’s ability to guarantee either the lineage of their training data or the feasibility of fine-tuning atop newer models trained on what are essentially summaries of data from older models, but time will tell. For now, it’s nice to see Meta pushing the more pragmatic parts of LLMs (i.e., availability and performance at small scale).