Live data from Hacker News

Google releases Gemma 4 open models

deepmind.google

391–400 of 507 posts

Re: Google releases Gemma 4 open models

#391

Earlier quoted context omitted.

I find it interesting that it decided to write a script despite not having access to tools, and is apparently aware of this lack of access since it then proceeds to do the computation manually. It's impressive it got as close as it did with estimates (and that it can actually do basic math now). Yet then it goes "let's do a precise check using tools" and just blatantly makes the whole thing up. Comedic but also worri…

Agree on all points! The difference between Gemma and Qwen here is that Qwen followed a much more detailed process - it consider leap years and seconds in its calculations (where Gemma used estimates like "roughly x years")

Turns out I wasn't reading closely enough. Notice that it first comes up with the number out of thin air prior to the math which is to "verify" it.

Following this charade, the "precise check" using "common tools" (which it does not have access to) pulls an entirely different number out of thin air.

It then asks if this new different number is correct, checks by "converting it back" with a utility it doesn't have access to, declares success, and then prints this second number.

Both numbers are wrong.

The fact that I was so easily misled on such a basic task when I was actively interested in where things had gone wrong is concerning to say the least. I'm beginning to think that thinking traces are actually quite nefarious in many contexts and that the entire exercise is some sort of trained hallucination task as opposed to even remotely resembling what's actually going on.

Re: Google releases Gemma 4 open models

#392

Thinking / reasoning + multimodal + tool calling. We made some quants at https://huggingface.co/collections/unsloth/gemma-4 for folks to run them - they work really well! Guide for those interested: https://unsloth.ai/docs/models/gemma-4 Also note to use temperature = 1.0, top_p = 0.95, top_k = 64 and the EOS is " ". " thought\n" is also used for the thinking trace!

and the EOS is " ". " thought\n" is also used for the thinking trace! Can someone explain this to me? Why is this faux-XML important here?

That’s how the model is trained to signal the end to its generation and to indicate its thinking.

Re: Google releases Gemma 4 open models

#393

Earlier quoted context omitted.

I gave it a whirl but was unenthused. I'll try it again, but so far have not really enjoyed any of the nvidia models, though they are best in class for execution speed.

I'll pipe in here as someone working on an agentic harness project using mastra as the harness. Nemotron3-super is, without question, my favorite model now for my agentic use cases. The closest model I would compare it to, in vibe and feel, is the Qwen family but this thing has an ability to hold attention through complicated (often noisy) agentic environments and I'm sometimes finding myself checking that i'm not on…

I’d be very interested in trying it if you could spare the time to write up how to tune it well. If not thanks for the input anyway.

Re: Google releases Gemma 4 open models

#394
post #320

Thinking / reasoning + multimodal + tool calling. We made some quants at https://huggingface.co/collections/unsloth/gemma-4 for folks to run them - they work really well! Guide for those interested: https://unsloth.ai/docs/models/gemma-4 Also note to use temperature = 1.0, top_p = 0.95, top_k = 64 and the EOS is " ". " thought\n" is also used for the thinking trace!

I'm trying to disable "thinking", but it doesn't seem to work (in llama.cpp). The usual `--reasoning-budget 0` doesn't seem to change it, nor `--chat-template-kwargs '{"enable_thinking":false}'` (both with `--jinja`). Am I missing something? EDIT: Ok, looks like there's yet another new flag for that in llama.cpp, and this one seems to work in this case: `--reasoning off`. FWIW, I'm doing some initial tries of unsloth…

Oh very cool! Will check the `--reasoning off` flag as well!

Yep the models are really good!

Re: Google releases Gemma 4 open models

#395

Thinking / reasoning + multimodal + tool calling. We made some quants at https://huggingface.co/collections/unsloth/gemma-4 for folks to run them - they work really well! Guide for those interested: https://unsloth.ai/docs/models/gemma-4 Also note to use temperature = 1.0, top_p = 0.95, top_k = 64 and the EOS is " ". " thought\n" is also used for the thinking trace!

Temperature 1.0 used to be bad for sampling. 0.7 was the better choice, and the difference in results were noticeable. You may want to experiment with this.

You might be right, but Google's recommendation was temp 1 etc primarily because all their benchmarks were used with these numbers, so it's better reproducibility for downstream tasks

Re: Google releases Gemma 4 open models

#396

Earlier quoted context omitted.

Temperature 1.0 used to be bad for sampling. 0.7 was the better choice, and the difference in results were noticeable. You may want to experiment with this.

You might be right, but Google's recommendation was temp 1 etc primarily because all their benchmarks were used with these numbers, so it's better reproducibility for downstream tasks

Fair, though putting a note in the readme about temperature 0.7 couldn't hurt.

I wonder why they do benchmarks with 1 instead of 0.7... that's strange. 0.7 or 0.8 at most gives noticeably better samples.

Re: Google releases Gemma 4 open models

#397

Thinking / reasoning + multimodal + tool calling. We made some quants at https://huggingface.co/collections/unsloth/gemma-4 for folks to run them - they work really well! Guide for those interested: https://unsloth.ai/docs/models/gemma-4 Also note to use temperature = 1.0, top_p = 0.95, top_k = 64 and the EOS is " ". " thought\n" is also used for the thinking trace!

Wow! Thank you very much!

Thanks!

Re: Google releases Gemma 4 open models

#398
post #274

Thinking / reasoning + multimodal + tool calling. We made some quants at https://huggingface.co/collections/unsloth/gemma-4 for folks to run them - they work really well! Guide for those interested: https://unsloth.ai/docs/models/gemma-4 Also note to use temperature = 1.0, top_p = 0.95, top_k = 64 and the EOS is " ". " thought\n" is also used for the thinking trace!

neat, time to update my spam filter model hehe

Haha! Ye the model is really good
Post reply on HN