[flagged]
Google releases Gemma 4 open models
471–480 of 507 posts
Re: Google releases Gemma 4 open models
#472Hi all! I work on the Gemma team, one of many as this one was a bigger effort given it was a mainline release. Happy to answer whatever questions I can
what part of gemma did you contribute to?
Re: Google releases Gemma 4 open models
#473Hi all! I work on the Gemma team, one of many as this one was a bigger effort given it was a mainline release. Happy to answer whatever questions I can
Is there going to be a new ShieldGemma based on Gemma 4?
Re: Google releases Gemma 4 open models
#474I ran these in LM Studio and got unrecognizable pelicans out of the 2B and 4B models and an outstanding pelican out of the 26b-a4b model - I think the best I've seen from a model that runs on my laptop. https://simonwillison.net/2026/Apr/2/gemma-4/ The gemma-4-31b model is completely broken for me - it just spits out "---\n" no matter what prompt I feed it. I got a pelican out of it via the AI Studio API hosted model…
It seems unreasonable to expect an LLM to have an accurate "mental model" of a bicycle since most humans don't either, and it's our written descriptions the LLM is learning from. A multi-modal model trained on captioned pictures isn't much better off, since what would induce it to memorize the details that we also abstract away ("a frame connecting it all together") ? Even posessing AGI, most humans still can't reason their way to a functional bicycle.
Comparing bicycles between LLMs doesn't really tell us much, since how do you differentiate an AI with a good model of a bicycle, but that does a poor job of drawing one with SVG, vs one that that has a much worse model but is in fact doing a great job of rendering it?!
I suppose you could say the same for the Pelican, although it does seem more reasonable to guess that most models could accurately describe the body plan of an animal even if they can't do a good job of drawing one with SVG.
Re: Google releases Gemma 4 open models
#475Thinking / 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!
llama.cpp (b8642) auto-fits ~200k context on this 24GB RX 7900 XTX & it shows a solid 100+ tok/s ("S_TG t/s") on the first 32k of it, nice! ./llama-batched-bench -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL \ -npp 1000,2000,4000,8000,16000,32000,64000,96000,128000 -ntg 128 -npl 1 -c 0 | PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s | |-------|--------|------|--------|----------|----------…
~50 tok/s on M1 Max 64Gb
Re: Google releases Gemma 4 open models
#476I ran these in LM Studio and got unrecognizable pelicans out of the 2B and 4B models and an outstanding pelican out of the 26b-a4b model - I think the best I've seen from a model that runs on my laptop. https://simonwillison.net/2026/Apr/2/gemma-4/ The gemma-4-31b model is completely broken for me - it just spits out "---\n" no matter what prompt I feed it. I got a pelican out of it via the AI Studio API hosted model…
It seems unreasonable to expect an LLM to have an accurate "mental model" of a bicycle since most humans don't either, and it's our written descriptions the LLM is learning from. A multi-modal model trained on captioned pictures isn't much better off, since what would induce it to memorize the details that we also abstract away ("a frame connecting it all together") ? Even posessing AGI, most humans still can't reaso…
For anyone who downvoted this due to thinking that humans, hence LLMs, DO have a good model of a bicycle, I challenge you to draw one.
No cheating and looking at pictures. Pen and paper. Do the easy bit first and draw wheels, seat, handlebars, pedals and chain. Add a stick figure riding it if that helps.
Now draw the frame.
Now google a photo of a bicycle.
Re: Google releases Gemma 4 open models
#477So what's the business strategy here? Google is the only USA based frontier lab releasing open models. I know they aren't doing it out of the goodness of their hearts.
Spreading propaganda through aligned model censored to eschew wrongthink? I mean, I truly believe there's some of that in the LLM world, but probably not the real reason you're searching for. Might be trying to (re)gain mindshare/cred amongst the hackers.
Re: Google releases Gemma 4 open models
#478Gemma-4-E4B-it scored 15/25 on my https://sql-benchmark.nicklothian.com/#all-data (agentic SQL generation). The naming is a bit odd - E4B is "4.5B effective, 8B with embeddings", so despite the name it is probably best compared with the 8B/9B class models and is competitive with them. Qwen3.5-9B also scores 15/25 in thinking mode for example. The best 9B model I've found is Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distil…
Very happy to see updates to your benchmark. Looking forward to inclusion of larger Gemma 4 models!
Re: Google releases Gemma 4 open models
#479Earlier quoted context omitted.
Do you think it's just part of their training set now?
It's time to do "frog on a skateboard" now.
In case you haven't seen this: https://x.com/JeffDean/status/2024525132266688757
Re: Google releases Gemma 4 open models
#480Earlier quoted context omitted.
Daniel, your work is changing the world. More power to you. I setup a pipeline for inference with OCR, full text search, embedding and summarization of land records dating back 1800s. All powered by the GGUF's you generate and llama.cpp. People are so excited that they can now search the records in multiple languages that a 1 minute wait to process the document seems nothing. Thank you!
Oh appreciate it! Oh nice! That sounds fantastic! I hope Gemma-4 will make it even better! The small ones 2B and 4B are shockingly good haha!
Just switched from 3.1 Flash Lite to Gemma-4 31B on the AI Studio API since there is a generous 1500/day on non-billed projects. It's doing fantastic.