Gemma Developer Day Paris 2025
2025-03-12
Paris, France
who it's for
ml engineers, applied machine learning engineers, computer vision engineers, mobile engineers, robotics engineers, safety and security engineers, infrastructure and ML ops engineers
topics
Gemma 3 architecture and checkpoints; 128k long-context tokenizer; multilingual pretraining (140 languages); q4_0 and FP8 quantization; on-device and mobile inference best practices; SigLIP vision encoder integration; image safety classification (ShieldGemma); text watermarking with SynthID; RLHF ordering-based reward techniques; agent orchestration and function calling; fine-tuning workflows and LoRA adapters; deploying Gemma on Jetson/Ollama/Transformers
| title | author | length |
|---|---|---|
| Welcome to Gemma 3 | Tris Warkentin | 12m |
Gemma 3 launches today as a faster, more capable Gemma family model built on Gemini 2.0 technology and tuned to run from phones to workstations. It claims… | ||
| Deepdive into Gemma 3 | Aishwarya Kamath, Cassidy Hardin, Shreya Pathak, Robert Dadashi, Léonard Hussenot, Geoffrey Cideron | 27m |
Gemma 3 ships as four open checkpoints (1B, 4B, 12B, 27B) with pre/post-train weights and quantizations bfloat16, float8, 4, and q4_0; the 1B in q4_0 fits… | ||
| Building intelligent agents with Gemma 3 | Kathleen Kenealy | 10m |
Shows how to assemble an agent using Gemma models with concrete parts: a reasoning model that plans and decomposes goals, a tool layer supporting function… | ||
| Gemma 3 architecture and design | Alek Andreev | 10m |
Gemma 3 uses a 256,000-token tokenizer updated for 64 languages, improving multilingual and finetune coverage; models come in 1B, 2B, 4B, 12B, 27B with… | ||
| A truly multilingual Gemma 3 | Adi Mayrav Gilady | 09m |
Gemma 3 claims state‑of‑the‑art multilingual fluency by pretraining on 140 languages in pre-training and providing over 35 languages out of the box after… | ||
| Leveraging large-scale models for efficient learning | Morgane Rivière | 08m |
Large pretrained models guide smaller ones by matching the teacher’s next-token probability distributions rather than hard one-hot targets, which speeds… | ||
| RLHF for finer alignment with Gemma 3 | Louis Rouillard | 10m |
Gemma 3 uses two concrete fixes to reduce RLHF reward hacking: train the reward model from the same policy to keep RM judgments in-distribution and build an… | ||
| Open models and open compression | Edouard YVINEC | 09m |
Claims that Gemma runs compressed without quality loss, showing Q4_0 (≈4.5 bits per weight, one scale per 32 values) and an FP8 variant about 2× smaller than… | ||
| PaliGemma – Making Gemma 2 see by adding a vision encoder | Andreas Steiner | 11m |
PaliGemma adds a SigLIP vision encoder to Gemma 2, converting images to tokens for the LLM and enabling zero-shot segmentation and detection by encoding… | ||
| ShieldGemma 2 – Developing safe and responsible AI for images | Dana Kurniawan, Wenjun Zeng | 07m |
ShieldGemma 2 is a 4 billion parameter image safety classifier built on Gemma 3, trained on curated synthetic and natural images for sexually explicit, violent… | ||
| SynthID – Watermarking and identifying AI-generated text | Tatiana Matejovicova | 08m |
SynthID-Text implements a lightweight watermark that embeds a secret key into generated text by computing per‑n‑gram bits and using two‑token tournament… | ||
| Using and fine-tuning Gemma 3 | Etienne Pot | 09m |
Gemma 3 ships a Gemma Python library that provides chat, multi‑turn caching, and multimodal prompts with spatial tokens, and it integrates with Kauldron… | ||
| Gemma on mobile and web. Best and worst practices | Mark Sherwood | 14m |
Mobile and web deployment focuses on concrete tradeoffs and tools not fluffy promises: use Gemma 3 1B quantized to 529 megabytes (zipped under 400 MB) to run… | ||
| Welcome to the Gemmaverse | Omar Sanseviero | 07m |
The Gemma ecosystem reports over 60,000 projects on Hugging Face and ships Gemma 3 with concrete upgrades: 128k context length, a 1B model plus… | ||
| Panel: The importance of open models | Marzieh Fadaee, Laurent Mazare, Tris Warkentin, Gus Martins | 28m |
Panel details why open models matter: Gemma’s technical report reveals KV‑cache mechanics and a pretrain that supports 140 languages, Cohere’s Aya family… | ||
| Live demo: Gemma 3 on Jetson Orin Nano | Asier Arranz | 18m |
Jetson Orin Nano boots Jetpack, flashes via SDK Manager, and runs Gemma 3 (demo uses a Gemma 3 4B build served with Ollama) after increasing swap to 16 GB… | ||
| Running Gemma on Mac and Windows PCs with Ollama | Michael Chiang, Jeffrey Morgan | 09m |
Demonstration shows Gemma 3 on Ollama running locally and in serverless Google Cloud Run, with Ollama claiming ~150 million model downloads and tens of… | ||
| Integrating the Gemma model family in Transformers | Arthur Zucker | 15m |
Finished integrating Gemma 3 in Transformers and shipping device‑agnostic code with integration tests, CLI support and Hugging Face repo weights for… | ||
| LeRobot – Lowering the entry barrier to AI for robotics | Remi Cadene | 15m |
LeRobot is an open-source Python library from Hugging Face that bundles pretrained models, datasets, simulators and robot middleware and claims to let novices… | ||
| nothing matches | ||