It is interesting, this model, but not as good as they claim, in my limited testing.
I have a small SQL test puzzle that it dismally fails to solve even though it is a commonplace student puzzle that is clearly in the training set because even when prompted towards the solution it refutes that it will work.
Gemma 4 E2B does the same (but is not a thinking model). Gemma 4 E4B can at least be prompted to offer up the solution with additional hints and serious suggestions.
I've not tried the Deepseek model in the (IIRC) 9B range get.
Gemma 4 12B in thinking mode jumps in and solves my problem immediately.
It is interesting, this model, but not as good as they claim, in my limited testing.
I have a small SQL test puzzle that it dismally fails to solve even though it is a commonplace student puzzle that is clearly in the training set because even when prompted towards the solution it refutes that it will work.
Gemma 4 E2B does the same (but is not a thinking model). Gemma 4 E4B can at least be prompted to offer up the solution with additional hints and serious suggestions.
I've not tried the Deepseek model in the (IIRC) 9B range get.
Gemma 4 12B in thinking mode jumps in and solves my problem immediately.