Google’s Gemma 3: The Best Open-Weight Model Yet?

15 Mar 2025 · 13 min
12:47

This is a transcript of a YouTube video by LARRUZ reviewing Google’s Gemma 3 model family. The speaker summarizes that Gemma 3 is a notable open-weight family of four models (1 billion → 27 billion parameters) and highlights its strong human-preference performance (Chatbot Arena top-10 entry with a score quoted as 1339). Larger Gemma 3 models are multilingual (claimed “>140 languages”) and multimodal (text, images, short video); the 1B model is English-only and has a smaller context window. Pretraining token counts differ across sizes (speaker cites ~2 trillion tokens for the smaller model and ~14 trillion for the largest). Google used a four-stage post-training pipeline: distillation from larger instruct models, RLHF (human feedback), reinforcement from machine feedback (to improve math), and reinforcement from execution feedback (to improve coding). Weights (base + instruct) and official quantized releases (32/16/8/4-bit) are available on Hugging Face; the speaker also summarizes hardware guidance (mentions single H with 80 GB VRAM for useful 4-bit runs and consumer GPUs like “3090 1490 or 5090” for other quantized runs). The reviewer finds Gemma 3 strong for creative writing and chat, less recommended for coding on smaller models, and notes some multimodal/demo limitations during their tests.

Questions this video answers

1339 (quoted as entering the Chatbot Arena top-10 with score 1339).

Four models ranging from 1 billion to 27 billion parameters (speaker: "bei einer Milliarde beginnen und bis zu 27 Milliarden reichen").

More than 140 languages (speaker: "mehr als 140 Sprachen").

32-bit, 16-bit, 8-bit, and 4-bit quantized versions (speaker lists 32 bit, 16 bit, 8 bit and 4 bit).

On Hugging Face (speaker: "Alle Gewichtungen für alle vier dieser Modelle sind auf Hugging Face verfügbar").

Chapters

More from LARRUZ

8:49
youtube.com

OpenAI’s New AI Is Doing Things We Barely Understand

3d ago
5:38
youtube.com

DeepSeek Solved AI’s Memory Problem

5d ago
33:17
youtube.com

DeepMind's New AI Just Cracked The Code Of Life

1w ago
3:43
youtube.com

The Billion Dollar AI Advantage Is Disappearing

1w ago
5:14
youtube.com

Claude Opus 5.5 AI: An Incredible Leap Forward

3w ago
5:16
youtube.com

Yes, Jev Is Insane, But There's A Catch

3w ago
5:11
youtube.com

DeepSeek’s Insane New Architecture

4w ago
4:16
youtube.com

Claude Is Now Leaving Invisible Fingerprints In Its Text

1mo ago
9:18
youtube.com

I Never Thought I’d See This Happen

1mo ago
5:22
youtube.com

GPT-6 Astra Changes Everything

1mo ago
4:55
youtube.com

Claude Fable AI Is Much Stranger Than The Headlines Suggest

1mo ago
See all content