3 ms·
The benchmark comparisons to Gemma 3 27B on Hugging Face are interesting: The Gemma 4 E4B variant (https://huggingface.co/google/gemma-4-E4B-it https://huggingf
by minimaxir 7mo ago
The benchmark comparisons to Gemma 3 27B on Hugging Face are interesting: The Gemma 4 E4B variant (https://huggingface.co/google/gemma-4-E4B-it https://huggingface.co/google/gemma-4-E4B-it) beats the old 27B in every benchmark at a fraction of parameters.
The E2B/E4B models also support voice input, which is rare.
- regularfry 7mo agoThinking vs non-thinking. There'll be a token cost there. But still fairly remarkable!
- DoctorOetker 7mo agoIs there a reason we can't use thinking completions to train non-thinking? i.e. gradient descent towards what thinking would have answered?
- joshred 7mo agoFrom what I've read, that's already part of their training. They are scored based on each step of their reasoning and not just their solution. I don't know if it's still the case, but for the early reasoning models, the "reasoning" output was more of a GUI feature to entertain the user than an actual explanation of the steps being followed.