Vitalik Buterin thinks local AI is finally getting close. Not perfect. But close.
On September 17, the Ethereum co-founder posted about how recent improvements in Qwen 3.8 Flash and llama.cpp have pushed on-device AI to a point where it can handle a meaningful range of tasks on his Strix Halo laptop — without shipping personal data to a remote server. For anyone watching the intersection of AI and crypto wallets, it’s a pretty significant shift in what’s actually possible on consumer hardware.
The benchmark numbers he shared are worth sitting with. Input processing ran between 109.82 and 373.22 tokens per second. Output generation came in at 18.42 to 33.37 tokens per second. Those aren’t theoretical lab figures — they’re from Buterin’s own machine, running locally. And for a setup that keeps your data off the cloud, that’s fast.
From April to September: What Changed
Back in April, Buterin’s picture of local AI was narrower. Models like Qwen3.5:35B were basically handling light work — transcription, summarization, that kind of thing. The heavier lifting still went to remote systems. It wasn’t a privacy-first setup so much as a privacy-adjacent one.
By September, that changed. Local models had grown capable enough to act as the primary interface for a broader set of tasks. Remote models don’t disappear from the picture entirely — they’re still there — but they only get called in when necessary. The key move is controlling which data actually leaves the device. That’s where the privacy gain lives.
It’s a meaningful distinction. Most AI setups today assume a constant connection to a cloud backend. Buterin’s framing flips that default: local first, remote only when the task demands it. For users who don’t want their activity logged or analyzed by a third-party server, that matters.
Qwen3.8-Flash-Next and the Wallet Problem
The model at the center of this is Qwen3.8-Flash-Next, built by Alibaba’s Qwen team. It runs on 125 billion parameters but manages to do it with fewer resources than earlier models by activating only a fraction of those parameters per token. There are also additional n-gram embedding tables in the architecture. The result is a model that punches above its weight on efficiency — which is exactly what you need if you’re running it locally on a laptop rather than a data center.
But here’s where it gets complicated for crypto. Efficiency and speed don’t automatically translate to trustworthiness in a financial context. Buterin flagged three specific concerns that haven’t been solved yet: model judgment, resistance to malicious instructions, and transaction authorization. Those aren’t minor footnotes. They’re basically the whole ballgame when you’re talking about a system that could theoretically interact with a crypto wallet.
Prompt injection is probably the nastiest of those problems. If a bad actor can slip malicious instructions into data the model processes — a webpage, a document, an email — the model might act on those instructions without the user realizing it. That’s a serious attack surface for anything touching financial assets.
So the practical answer, at least for now, is clear: low-risk automations can run within strict permissions, but anything involving wallet transactions still needs a separate human confirmation step. Deterministic policies, not AI discretion, govern what actually moves money.
Steward, EIP-7906, and What’s Still Pending
The Ethereum ecosystem isn’t just watching this from the sidelines. There’s active work on integrating local AI concepts into wallet infrastructure. Steward, a macOS smart-account wallet, is one of the projects testing these ideas in practice. It’s not fully deployed yet, and audit status is still pending — so it’s more of a proof-of-concept than a shipping product right now.
And then there’s EIP-7906. The draft proposal puts forward a framework built around post-transaction assertions. The idea is that after a transaction executes, the system checks the outcome against a set of predefined conditions. If something doesn’t match — if the account state changed in a way that wasn’t supposed to happen — the mechanism can catch it. It’s a way to build a safety net underneath AI-assisted transactions without relying purely on the AI’s own judgment to stay in bounds.
That’s a smart architectural choice. It doesn’t try to make the AI perfectly trustworthy. It assumes the AI might sometimes get it wrong and builds a verification layer on top. Whether that’s enough for mainstream wallet users is unclear yet.
What’s not unclear is the direction. Buterin’s September update isn’t a product announcement. It’s a progress report from someone who’s been tracking these capabilities closely and thinks the gap between “interesting experiment” and “practically useful” is closing faster than most people expected.
373 tokens per second on a laptop. 125 billion parameters, fewer resources than last year’s models. A draft EIP that could let smart accounts verify their own transaction outcomes. None of it is production-ready for high-stakes wallet use. But it’s not theoretical anymore either.
Frequently Asked Questions
What token-per-second speeds did Buterin’s local AI setup reach?
Input processing ran between 109.82 and 373.22 tokens per second, while output generation came in at 18.42 to 33.37 tokens per second on his Strix Halo laptop.
What is EIP-7906 and how does it relate to AI wallet security?
EIP-7906 is a draft proposal that uses post-transaction assertions to check whether a transaction’s outcome matches predefined conditions, adding a verification layer for AI-assisted crypto wallet activity.
