Ethereum co-founder Vitalik Buterin posted an updated assessment of local AI capabilities on September 17, arguing that Qwen 3.8 Flash running on a Strix Halo laptop can handle a large share of tasks privately while withholding user context from remote systems.
The shift marks a change from Buterin’s April position, when he described earlier Qwen models as limited to bounded tasks and familiar programming work. In the intervening months, local inference performance has improved enough that Buterin now sees practical application in crypto wallets, though with strict guardrails.
Performance metrics and model specifications
Qwen 3.8 Flash is an open-weight multimodal mixture-of-experts model released by Alibaba’s Qwen team. The model has 125 billion parameters total, with a 51 billion parameter n-gram embedding table, but only 6 billion parameters activate per token. Across 10 workloads, the model achieved input-processing rates of 109.82 to 373.22 tokens per second, and output generation rates of 18.42 to 33.37 tokens per second on the Strix Halo laptop.
In capability benchmarks, Qwen evaluated Flash-Next against 14 tests. On 8 of those benchmarks, the Flash-Next model outperformed Qwen3.7-Plus, a larger variant, suggesting efficiency gains in the newer architecture.
Wallet security architecture
Buterin’s framework for AI-assisted wallets goes beyond model judgment alone. According to his April guidance, wallet security requires fixed policies, deterministic transaction limits, and a human-plus-model approval rule for risky actions. He emphasized that policy enforcement must sit outside the language model itself, creating a human-confirmation firewall between user intent and transaction execution.
The Ethereum Foundation has funded Steward, a fully local macOS smart-account wallet designed to run on-device with a light client and AI assistant. Steward represents an attempt to operationalize Buterin’s wallet security principles in a production wallet.
Draft standards for transaction validation
EIP-7906, a draft proposal, introduces post-transaction assertion frames that inspect final state differences and reject outcomes violating specified conditions. This mechanism would allow wallets to encode user-defined constraints that execute after a transaction settles, adding a second layer of protection beyond pre-approval controls.
Buterin’s September post reflects a narrowing of the gap between local and remote model capabilities over six months. The practical implication is that privacy-sensitive wallet operations can remain on-device, delegating only computationally harder tasks to remote systems while keeping user context local.
