491–500 of 507 posts
Hi 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
Good work, it's quite close to Gemini 3 Pro in my tests, but 10x cheaper: https://aibenchy.com/compare/google-gemma-4-31b-it-medium/go...
I 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…
Same experience on the 31B - something’s wrong. The MoE works as expected though.
Earlier 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!
> your work is changing the world I realize this may have been hyperbole, but it sure isn't changing the world.
In the world of local models, Unsloth is one of the most significant projects there is.
Earlier quoted context omitted.
If you want the model to have function calls available you need to run it in an agentic harness that can do the proper sandboxing etc. to keep things safe and provide the spec and syntax in your system prompt. This is true of any model: AI inference on its own can only involve guessing, not exact compute.
Thanks, I am very new to this and just run models in LMStudio. I think it would be very useful to have a system prompt telling the model to run python scripts to calculate things LLMs are particularly bad at and run those scripts. Can you recommend a harness that you like to use? I suppose safety of these solutions is its own can of worms, but I am willing to try it.
Earlier quoted context omitted.
Here's Gemma's. Qwen's is a bit too long to post here. Input format: ISO 8601 string (2026-04-01T16:00:00Z). Target format: Unix timestamp (seconds since Jan 1, 1970). Year: 2026 Month: 04 (April) Day: 01 Time: 16:00:00 Timezone: Z (UTC/Zulu) Option A: Manual calculation (too error-prone). Option B: Use a programming language (Python/JavaScript). Option C: Knowledge of specific dates. Let's use Python logic internall…
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…
Earlier quoted context omitted.
Did you try it?
yes, I've now I tried both the 20 GB version (gemma4:31b) which is the largest on the page[1], and the ~10 GB version (gemma4:e4b). The 20 GB version was rather slow even when fully loaded and with some RAM still left free, and the 10 GB version was speedy. I installed openclaw but couldn't get it to act as an agent the way Claude Code does. If you'd like to see a video of how both of them perform with almost nothing…
Also, I think I can see some swap being used. The way to see if a model is loaded completely in ollama is to use ollama ps to check the output. If it starts hitting limits you'll see the split there and a unified memory box will start to swap. Along with the performance crashing down, of course.
Thanks for the video and results, though. Just hopefully constructive tips.
Earlier quoted context omitted.
yes, I've now I tried both the 20 GB version (gemma4:31b) which is the largest on the page[1], and the ~10 GB version (gemma4:e4b). The 20 GB version was rather slow even when fully loaded and with some RAM still left free, and the 10 GB version was speedy. I installed openclaw but couldn't get it to act as an agent the way Claude Code does. If you'd like to see a video of how both of them perform with almost nothing…
The massive black borders are making the actual part of the video hard to see. Recording just the window and/or zooming the text as big as you can would make it work better. Also, I think I can see some swap being used. The way to see if a model is loaded completely in ollama is to use ollama ps to check the output. If it starts hitting limits you'll see the split there and a unified memory box will start to swap. Al…
Regarding the black borders, I've cropped, re-encoded this and reuploaded this as 1080p (the resolution the headless Mac gave over VNC) so you can watch that version without any black borders if you want: https://www.youtube.com/watch?v=5VOiH2zjAss
(not sure how large your screen is but this should be full size if you maximize it I guess). It's a re-encoding so it doesn't look as good as the original but you should be able to read anything you were interested in seeing. Next time I'll be sure to zoom in on the text more.
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 updated Ollama (again) and changed my windows swap file settings to use up to 200 Gb of C: (an SSD). On the largest model (gemma4:31b), I seem to be getting about 5 tokens per second. This is amazing to me, because I'm using a $100 computer, without any fancy GPU. I love watching it "think". Consider this is thousands of times faster than any written conversations in the past. Those involved pieces of paper being t…
But generally, I'd like to see above 20, >50 is mostly great, and more is better. For conversational response, that is, not batch or interactive loop.