DeepMind's Gemma 4 runs on phones and a Nintendo Switch
Two Minute Papers covers Google DeepMind's Gemma 4, a free and open family of models whose smallest variants run on phones and even a first-generation Nintendo Switch without an internet connection.
Score breakdown
Developers building agentic or AI-assisted apps can deploy Gemma 4 locally — on phones or low-end hardware — eliminating cloud dependency and subscription risk entirely.
- 01Gemma 4 is a free and open family of models released by Google DeepMind.
- 02The smallest Gemma 4 variants require only a few gigabytes of memory and run on phones without an internet connection.
- 03The 2 billion parameter Gemma 4 model runs on a first-generation Nintendo Switch.
Two Minute Papers, hosted by Dr. Károly Zsolnai-Fehér, frames Google DeepMind's Gemma 4 release as a direct answer to the risks of relying on proprietary AI subscriptions — citing reports of some Claude users losing access due to "heavy workloads" as a motivating example. Gemma 4 is described as a free and open family of models where the smallest variants require only a few gigabytes of memory, making them runnable on consumer phones without an internet connection and even on a first-generation Nintendo Switch. Within days of release, community members had already built offline translation apps, summarization tools, and real-time in-browser image classification demos, and fine-tuning work was already publicly available.
The strong performance of a dense model at this scale is presented as a genuine surprise.
The video highlights four surprising findings about Gemma 4. Most notably, the larger 31B parameter model is a dense model — meaning it activates all parameters at inference time — yet it achieved the #3 ranking among open models and outperformed some models 10 times its size, remaining competitive with some 20 times its size on certain benchmarks. This is contrasted with the dominant mixture-of-experts (MoE) architecture used by many modern large models, which routes inputs to only a subset of specialized sub-networks (typically 2 to 8 experts) to keep inference costs manageable. The strong performance of a dense model at this scale is presented as a genuine surprise.
Key facts
- 01Gemma 4 is a free and open family of models released by Google DeepMind.
- 02The smallest Gemma 4 variants require only a few gigabytes of memory and run on phones without an internet connection.
- 03The 2 billion parameter Gemma 4 model runs on a first-generation Nintendo Switch.
- 04Community members built offline translation, summarization, and real-time browser-based image classification apps within days of release.
- 05The 31B Gemma 4 model ranked #3 among open models and beat some models 10 times its size.
- 06The 31B model is a dense model, not a mixture-of-experts (MoE) architecture.
- 07Fine-tuning support for Gemma 4 was already publicly available shortly after release.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Apr 22, 2026 · 11:07 UTC. How this works →