[dannyw]

I'm confident we will continue to see improvements in edge models at minimum.

For example, Google has a vested interest in making Gemma as good as it can, because ultimately any edge inference is free for Google and they have a massive install base.

It wouldn't surprise me if Apple eventually trained their own foundational models, and while it'd be surprising, I wouldn't be shocked if Apple also released open weight edge models one day. Their Mac business is benefiting a lot from local AI, and for an extremely long list of reasons, Apple doesn't want either OpenAI or Anthropic to "win" and supersede their walled gardens as the first entry point.

There are a lot of incentives for open weight, Mistral is probably going to keep on making at least some open models, and meanwhile in the image generation space, we've got Krea 2, Ideogram 4, and BFL continues to ship most of their models as weights available (but under non-commercial licenses).

Oooh, and NVIDIA has been releasing open source ML, and now LLM models for what, a decade? NVIDIA likes selling hardware, Nemotron models, while rarely topping leaderboards (probably not benchmaxxed), are generally fine and really great models for fine-tuning or CPT.

[torginus]

'Open weight' is roughly analogous to 'freeware' - software that's free to use, but closed source. We haven't seen a third category - 'paid weight' - software where you pay for the weights, but you keep them and have to run them on your own machines - analogously, updates could so if you don't want your LLMs cutoff to stay at a point forever, you have to keep buying new weights. These are standard practices in software pricing.

[Alpha3031]

Apple does have its own models for on-device use and the newest series has a MoE with expert activation swapped per prompt instead of per token which is interesting.

I'd assume it's probably somewhat based off Gemini distillation or something, and I'm not expecting them to release the weights this time at least, but it would definitely be interesting to see an open local model using a similar MoE whether that's at a similar size (20B A1–4B supposedly) or a larger model (something DeepSeek V4 Flash sized, or slightly smaller, would just barely fit on consumer hardware, and could be fun to see how it performs I think).