Six Yale computer scientists have described Maverick, a protocol for running an open-source AI model mostly on someone else’s machine: the user’s phone or other low-powered device passes the heavy computing to an outside server, which receives only a scrambled copy of the user’s input, and checks the result when it comes back. The paper went up on arXiv on Sept. 9, 2026, as a preprint; it has not been accepted at a conference or journal and does not link to code.
Checking that a machine you don’t control computed honestly is a problem zero-knowledge proof systems and some decentralized compute networks also work on. Maverick also keeps the user’s input hidden from that machine. DeepProve, a zero-knowledge machine learning system, checks results but leaves the input exposed to the server, according to the paper. Two authors, assistant professor Fan Zhang and PhD student Wenhao Wang, list the Initiative for CryptoCurrencies and Contracts (IC3) as an affiliation, and an IC3 grant supports part of Wang’s research.
Discussing blockchain marketplaces that sell AI inference, the running of trained models, host Kain Warwick described the problem of trusting someone else’s machine this month: “people are selling you inference, and you don’t really know where it’s coming from,” he said on Unchained’s Uneasy Money.
Get Unchained’s crypto news in your inbox with the free Unchained Daily newsletter.
How Maverick Works
Maverick sends most of the model’s math to the server, after the user’s device hides its input under a random-looking mask that it later strips from the answer. To check the server, the device runs a quick randomized test instead of redoing the work.
Co-author Charalampos Papamanthou, listed on Maverick as Yale only, is chief scientist at Lagrange Labs, according to his Yale page. Lagrange builds zero-knowledge proof systems, including DeepProve, and Papamanthou co-wrote the DeepProve paper.
What the Tests Show
In the paper’s end-to-end tests, the biggest gain came when the user’s device prepared its masks ahead of time: prompt tokens went through up to 45 times faster than if that device had run the model by itself. Building masks during the run cut the peak to 17 times. Those figures come from one processing thread on the user’s device working with a CPU server allowed as many as 128 threads, an eight-token prompt and no network lag counted. In another test, which timed one short request over a simulated network with two device threads and 64 server threads, the two private modes returned their first output only 1.42 and 1.46 times faster than running locally at a 50-millisecond round trip, and slower than local at 100 milliseconds. Qwen3-4B was the only model tested.
The authors call Maverick’s client storage needs “a key bottleneck of our approach.”
Related Listen: Why Blockchains Could Undercut AI Labs on the Price of Inference
