Every figure, name and capability the finished video states, chased to a primary source.
Gemma 3n natively supports image, audio, video and text inputs, and text output.
“Gemma 3n natively supports image, audio, video, and text inputs and text outputs.”
Source: Google Developers Blog, Introducing Gemma 3n: The developer guide https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
This is the claim the whole video rests on, and it is worth being precise about what it does and does not say. The inputs are multimodal. The output is text. Nothing in Google’s own description says the model produces audio.
E2B and E4B, with raw parameter counts of 5B and 8B, running in roughly 2GB and 3GB of memory.
“While their raw parameter count is 5B and 8B respectively, architectural innovations allow them to run with a memory footprint comparable to traditional 2B and 4B models, operating with as little as 2GB (E2B) and 3GB (E4B) of memory.”
Source: Google Developers Blog, Introducing Gemma 3n: The developer guide https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
Google’s own model documentation puts the same idea a second way: E2B loads over 5 billion parameters in standard execution but carries “an effective memory load of just under 2 billion (1.91B) parameters.”
Source: Google AI for Developers, Gemma 3n model overview https://ai.google.dev/gemma/docs/gemma-3n
140 languages for text, and multimodal understanding of 35.
“Gemma 3n delivers quality improvements across multilinguality (supporting 140 languages for text and multimodal understanding of 35 languages)”
Source: Google Developers Blog, Introducing Gemma 3n: The developer guide https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
The two numbers are not the same number and the video keeps them apart. Text support is much wider than multimodal support.
MobileNet-V5.
“High-performance MobileNet-V5 encoder”
Source: Google AI for Developers, Gemma 3n model overview https://ai.google.dev/gemma/docs/gemma-3n
The developer guide adds that it “processes up to 60 frames per second on a Google Pixel” and “natively supports resolutions of 256x256, 512x512, and 768x768 pixels.”
Source: Google Developers Blog, Introducing Gemma 3n: The developer guide https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
32K tokens.
Source: Google AI for Developers, Gemma 3n model overview https://ai.google.dev/gemma/docs/gemma-3n
The video’s central correction is that the model writes text and something else turns that text into sound. Both named engines are real, separately maintained projects that run locally and are not part of Gemma 3n.
82 million parameters, Apache-2.0, 54 voices across 8 languages, 24 kHz output.
Source: hexgrad/Kokoro-82M model card, Hugging Face https://huggingface.co/hexgrad/Kokoro-82M
The model card describes it as an open-weight text-to-speech model that “delivers comparable quality to larger models while being significantly faster and more cost-efficient.” Published 27 January 2025.
A fast, local neural text-to-speech engine that embeds espeak-ng for phonemization. GPL-3.0.
“A fast and local neural text-to-speech engine that embeds espeak-ng for phonemization.”
Source: OHF-Voice/piper1-gpl https://github.com/OHF-Voice/piper1-gpl
Ollama’s own library documents Gemma 3 as multimodal, processing text and images, with a 128K context window and support for over 140 languages, requiring Ollama 0.6 or later.
Source: Ollama model library, gemma3 https://ollama.com/library/gemma3
This is the practical evidence for the video’s claim that taking images into a local model is something you can do now rather than something that is coming.