The part Google left out of the tweet:
> Gemma 4 comes in 4 sizes. The biggest (31B) just ranked #3 in the entire world on Arena AI. That's where real users vote on which AI gives better answers. Out of every open model on earth. Third.
The 26B model is even more interesting.
It uses something called "Mixture of Experts."
In English: 26 billion parameters total, but only 4 billion activate per task. Like having 26 employees but only 4 show up to any meeting. That's why it runs on a regular laptop despite being massive.
How smart is it?
89.2% on AIME 2026. AIME is one of the hardest math competitions on the planet. Most humans fail it. This model nearly aced it.
Codeforces ELO of 2,150. That's competitive programmer territory. It writes code better than 95% of devs on the platform ๐
256K context window = it reads and understands roughly 200 pages of text at once.
The smallest version (2.3B effective) runs on a Raspberry Pi. The 26B runs in your browser. No server. No cloud. Your browser tab, via WebGPU.
And the license change matters more than the benchmarks.
Previous Gemma versions had usage restrictions on what you could build. Gemma 4 ships under Apache 2.0 for the first time. Build anything. Ship commercially. Modify the weights. Zero strings attached.
Google just gave away one of the best AI models. For free.
App Maintenance Notice
To support mining reduction upgrades and system optimization, the app will undergo scheduled maintenance on February 17, 2026, 00:00โ12:00 UTC. Services will resume ASAP. Please follow official announcements for updates.
Network Updates (1/1/2026)
Active devices: 96,511
Top 3 tiers: Tier 10, Tier 9, and Tier 8
Most protocol fees are currently generated from Recast.
The holiday season has been a huge win for Puffpaw.
Wen 100k?
2,785,138,146 preVAPE is burned in protocol fees - 2.78% of total supply - accumulated from our network activities as of Dec.31 2025.
More details regarding the current network status will be shared in the next post.