Coinbase recently trained a smaller AI model for fraud detection at a cost of less than $100, and it performed better than larger models on its own tests.
What caught my attention was what Acurast did next.
Acurast tested the same Qwen3.5-9B model on 16 GB Android phones, getting its first response in about five minutes after deployment.
The team also tested more than 100 phones working in parallel, showing how AI tasks can be distributed across devices instead of relying on traditional cloud servers.
The reported average cost was just 0.57 ACU per phone per day during testing.
I find this interesting because not every AI task needs expensive GPUs. Smaller models running across ordinary phones could handle fraud checks, scam detection and background processing.
The next thing I'd watch is whether Acurast can maintain reliable performance as real demand grows.
That's where this experiment could become something much bigger.
Source: https://acurast.com/blog/network-and-tech/self-hosted-llm-phone/
$ACU
What caught my attention was what Acurast did next.
Acurast tested the same Qwen3.5-9B model on 16 GB Android phones, getting its first response in about five minutes after deployment.
The team also tested more than 100 phones working in parallel, showing how AI tasks can be distributed across devices instead of relying on traditional cloud servers.
The reported average cost was just 0.57 ACU per phone per day during testing.
I find this interesting because not every AI task needs expensive GPUs. Smaller models running across ordinary phones could handle fraud checks, scam detection and background processing.
The next thing I'd watch is whether Acurast can maintain reliable performance as real demand grows.
That's where this experiment could become something much bigger.
Source: https://acurast.com/blog/network-and-tech/self-hosted-llm-phone/
$ACU
