Coinbase did not beat Claude Opus with a bigger AI model. It used a smaller one.
Brian Armstrong shared that his team trained Qwen3.5-9B, a free open model, on Coinbase’s own fraud data.
It beat their Opus 4.5 reviewer on all four fraud tests. It also answered in less than half the time.
The training cost less than $100 on one Nvidia GPU.
So training smart AI is getting cheap. The next question is where to run it.
Acurast tested a surprising answer: a normal Android phone.
The model is only 5.7 GB after compression, so a 16 GB phone can hold it. It works with the same API style as OpenAI, so apps can connect easily.
Here is what the tests showed:
- Setup time: about 5 minutes
- Cost: roughly 5 to 10 US cents per phone per day
- Speed: up to 9.3 tokens per second on the fastest phones
It is not built for instant chat. Reading a very long prompt took over 3 minutes on one phone.
But it works well for background jobs. Each phone works on its own task, so 100 phones can handle 100 tasks at once. Think of checking flagged accounts or scanning for scams.
Your data also goes to your own setup, not to a big AI company.
I do not think phones will replace big GPU servers soon. But small specialist AI plus many phones working together is a new, cheaper way to run AI.
That is why I am watching @Acurast
$ACU
Brian Armstrong shared that his team trained Qwen3.5-9B, a free open model, on Coinbase’s own fraud data.
It beat their Opus 4.5 reviewer on all four fraud tests. It also answered in less than half the time.
The training cost less than $100 on one Nvidia GPU.
So training smart AI is getting cheap. The next question is where to run it.
Acurast tested a surprising answer: a normal Android phone.
The model is only 5.7 GB after compression, so a 16 GB phone can hold it. It works with the same API style as OpenAI, so apps can connect easily.
Here is what the tests showed:
- Setup time: about 5 minutes
- Cost: roughly 5 to 10 US cents per phone per day
- Speed: up to 9.3 tokens per second on the fastest phones
It is not built for instant chat. Reading a very long prompt took over 3 minutes on one phone.
But it works well for background jobs. Each phone works on its own task, so 100 phones can handle 100 tasks at once. Think of checking flagged accounts or scanning for scams.
Your data also goes to your own setup, not to a big AI company.
I do not think phones will replace big GPU servers soon. But small specialist AI plus many phones working together is a new, cheaper way to run AI.
That is why I am watching @Acurast
$ACU
