AI & Crypto

Yale scientists describe Maverick protocol for private AI inference

Yale researchers Fan Zhang and Wenhao Wang introduced Maverick, a protocol that allows users to run open-source AI models on remote servers using scrambled inputs. Testing with the Qwen3-4B model showed speed increases of up to 45 times when using pre-prepared masks.

Yale scientists describe Maverick protocol for private AI inference

Maverick protocol hides user inputs

Six Yale computer scientists described Maverick in a September 9, 2026, preprint on arXiv. The paper, which has not been accepted at a conference or journal and does not link to code, describes Maverick as a way to run an open-source AI model on someone else’s machine. The protocol allows a device to pass heavy computing to a server that receives only a scrambled copy of the input. To check the server, the device hides its input under a random-looking mask, strips it from the result, and runs a quick randomized test instead of redoing the work. This addresses the problem Kain Warwick described on Unchained regarding the difficulty of trusting a machine you do not control when people sell inference. 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 Wang’s research. Co-author Charalampos Papamanthou, listed as Yale only on Maverick, is the chief scientist at Lagrange Labs, which builds zero-knowledge proof systems like DeepProve. While DeepProve checks results, it leaves the input exposed to the server. Maverick provides a more private way to run open-source models.

Speed advantages and local bottlenecks

The researchers tested the protocol using the Qwen3-4B model with one processing thread on a device and up to 128 threads on a CPU server. These tests used an eight-token prompt and assumed no network lag.

Setup Speed relative to local
Pre-prepared masks Up to 45 times faster
Masks built during run 17 times faster
50ms round trip network 1.42 to 1.46 times faster

When the user device prepared masks ahead of time, prompt tokens moved up to 45 times faster than if the device ran the model by itself, but building masks during the run cut that speed to 17 times. In a test using two device threads and 64 server threads over a simulated network with a 50-millisecond round trip, the private modes returned output 1.42 to 1.46 times faster than running locally. This speed advantage disappears at a 100-millisecond round trip, where the protocol runs slower than local processing. The authors call Maverick client storage needs a bottleneck of the approach.