Open models in the Gemini era
summary
Google releases small open Gemma models (2B and 7B, pretrained and instruction tuned) built from Gemini research and trained on 6 trillion tokens, with Gemini offering 10 million token context windows. A demo shows Gemini 1.5 accurately identifies and localizes three web app bugs and proposes fixes. Gemma runs on CPU via Gemma.CPP and on H200 GPUs at tens of thousands of tokens per second, matches larger models on human evals, and supports LoRA fine‑tuning in Keras 3.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:00Open development in Google
- 1:56Google Brain and Google DeMind
- 2:50Google Products
- 3:40Future of AI
- 6:53Gemini
- 8:29Gemma
- 9:40Gemma goals
- 11:22Performance
- 12:06Progress
- 12:40Responsibility
- 13:10Openness
- 15:02Question
related talks
54mGemini, Google's Large Language Model
51mGemini, Google's Large Language Model
12mWelcome to Gemma 3
44mBuilding amazing Front-End Project with Gemma: A State-of-the-Art Open Model
40mBuilding amazing Front-End Project with Gemma: A State-of-the-Art Open Model
19mGemini 2.0: Beyond the Basics
15mGemma, DeepMind's Family of Open Models
18mIntroduction to Gemini APIs and AI Studio