Ethereum co-founder Vitalik Buterin said the latest wave of offline AI knowledge apps for phones is much better than the version he tried to build himself two months ago, though the tools still fall short of laptop-based models on difficult tasks.
In a Sept. 26 post on X, Buterin said the apps were still slower and less effective when questions became hard. He added that the weakest performance showed up in travel-focused queries. His test prompt asked for the best vegan restaurant in his current city, and he said every app he tried answered poorly.
Buterin said he wants these tools to keep improving and reach a point where users can easily look up real-world information without connecting to the internet.
A poidh bounty drew the submissions, with Buterin contributing 1 ETH
The link attached to his post pointed to bounty No. 31 on decentralized bounty platform poidh, titled “Build the best Android offline AI research app.” The name poidh comes from “pics or it didn’t happen.” On the platform, sponsors lock prize funds in a smart contract and participants submit proof of work to claim rewards.
According to BlockTempo’s reading of the poidh contract on Ethereum mainnet, the bounty was created on Sept. 18 by a community address. The description said it was an independent bounty inspired by Buterin’s earlier post, and that Buterin was not involved with the bounty itself and would not take part in judging unless he explicitly said so.
Still, the funding list shows Buterin’s public address, vitalik.eth, contributed 1 ETH. That accounted for more than 90% of the 1.098 ETH total prize pool. Using the article’s cited ETH price of about $2,684, the full bounty was worth roughly $2,947.
The idea followed a Sept. 17 post about a fully offline phone research tool
The bounty traces back to a Sept. 17 post from Buterin. He said he revisits the idea every few months. What he wants, he wrote, is a killer app that can run on a fully offline phone and deliver at least half the usefulness of “internet search plus frontier AI models,” because people often carry phones into places without internet access.
At that time, he said his own progress had only reached a 1 billion-parameter model running at about 10 tokens per second, and that it failed once the questions became interesting.
On model design, Buterin argued that phones are better suited to an extreme mixture-of-experts, or MoE, approach: around 100 billion total parameters, with most of them stored on disk, while fewer than 1 billion parameters are activated for each token. In that setup, parameters are split into many groups and only a small portion is called each time, so the rest can stay in storage instead of being loaded into phone memory.
Bounty rules set strict hardware and software limits
The bounty requires submissions to meet several conditions:
- The app must run on Android and GrapheneOS devices.
- Memory use is capped at 12GB.
- The combined size of the app, model and database cannot exceed 50GB.
- No network requests may be made after installation during use.
- The source code must be published on GitHub.
If Buterin publicly confirms that a submission meets the target, that project takes the entire 1.098 ETH prize. If no submission is confirmed by Oct. 31, the organizer and contributors will either receive refunds or choose a winner themselves.
Three submissions were on-chain as of Sept. 26
On-chain records show the bounty had received three submissions by Sept. 26, and none had been accepted. All three had already been submitted by the time Buterin posted, though he did not say which ones he tested.
Field Atlas
Field Atlas was submitted on Sept. 23. It uses the Qwen3 1.7B model with an offline knowledge pack, and its answers include cited sources.
Boar
Boar uses a Qwen2.5-1.5B small model with offline Wikipedia search. It can also add a data pack containing 50,000 entries and selects the smallest model capable of handling a given question based on difficulty.
AndroidLM
AndroidLM uses the Qwen3.6-35B-A3B model compressed to 2-bit precision. Its 12.3GB model file is streamed from storage and paired with a 21GB English Wikipedia database. For travel questions, users can also install a 0.3GB Wikivoyage guide. The developer tested it on a Pixel 8 Pro with 12GB of memory.
The model has 35 billion total parameters and activates about 3 billion per token. Among the three submissions, AndroidLM was the only one following the MoE direction Buterin had suggested. Even so, both its total parameter count and its active parameter count remain short of the 100 billion total and sub-1 billion active target he described.
No submission has been publicly confirmed by Buterin
As of now, Buterin has not publicly confirmed that any of the three entries meets the bounty standard. The deadline is Oct. 31.
His latest test results show clear progress in offline AI tools for phones. They also show the gap that remains in speed, difficult-query handling and real-world information retrieval. On the specific question of finding the best vegan restaurant in his current city, the current apps did not satisfy him.