WuBlockchain
Ethereum co-founder Vitalik Buterin has updated his work on local AI, noting that DeepSeek V4 now ships a 2-bit quantized version small enough to run inside roughly 90 GB of VRAM. On Apple hardware the model hits about 35 tokens per second; on AMD silicon, closer to 7.
Why it matters
Buterin said the real test for "CROPS AI" is multi-hardware support, not the decentralization pitch most projects lead with. A stack that only runs on a single accelerator class, or only at cloud scale, fails the local-compute promise even if it is technically distributed. Pushing the same quantized model across Apple and AMD is the early proof point that the architectural bet is portable.
He also laid out an "Ethereum access layer" that overlaps with CROPS AI: ZK-based paid remote LLM calls, private Ethereum RPC reads, and a push for more Ethereum-tuned AI models aimed at smart-contract and protocol-code security. The direction is to make the local model useful to Ethereum itself, not just a generic chatbot running on a developer laptop.
Market impact
The framing reframes the competitive landscape. "Decentralized AI" has been the dominant narrative across the sector; Buterin is signalling that the harder, more useful problem is hardware portability plus protocol-level integration. Watch for the next wave of CROPS-aligned projects to market themselves on Apple-Silicon and AMD support rather than node counts, and for Ethereum tooling — auditors, formal-verification pipelines, RPC providers — to start shipping CROPS-aware integrations.
Frequently asked questions
-
What did Vitalik Buterin say about CROPS AI in his latest update?
He said true "CROPS AI" should support multiple hardware platforms rather than being framed only as decentralized AI, and described a CROPS Ethereum access layer that includes ZK-based paid remote LLM calls and private Ethereum RPC reads.
-
How fast does the new 2-bit DeepSeek V4 model run locally?
The 2-bit quantized DeepSeek V4 build runs within roughly 90 GB of VRAM and reaches about 35 tokens per second on Apple hardware and around 7 tokens per second on AMD hardware.
-
What is the CROPS Ethereum access layer?
It is the part of Vitalik's CROPS AI proposal that overlaps with Ethereum itself: ZK-based paid remote LLM calls, private Ethereum RPC reads, and Ethereum-tuned AI models intended to improve smart-contract and protocol-code security.
-
Why does hardware portability matter for CROPS AI?
Because a stack that only runs on a single accelerator class or only at cloud scale fails the local-compute promise even if it is technically distributed. Supporting Apple and AMD silicon is the proof point that the architecture is portable.
-
How could this change what Ethereum AI projects focus on next?
Expect CROPS-aligned projects to market Apple-Silicon and AMD support over node counts, and for Ethereum tooling such as auditors, formal-verification pipelines, and RPC providers to begin shipping CROPS-aware integrations.
CoinTelegraph
WatcherGuru
Crypto News
CryptoSlate
CoinDesk