WuBlockchain
Ethereum co-founder Vitalik Buterin updated his progress on local AI, saying DeepSeek V4 now ships a 2-bit quantized version that runs in roughly 90 GB of VRAM, hitting about 35 tokens per second on Apple hardware and 7 tok/s on AMD silicon. He framed the milestone as the first concrete signal that CROPS-style private inference can run on consumer-grade machines, not just data-centre GPUs.
Why it matters
Vitalik pushed back on the "decentralized AI" label, arguing true CROPS AI should be hardware-agnostic — Apple, AMD, NVIDIA, and specialised accelerators all in. He also outlined a CROPS Ethereum access layer that overlaps with CROPS AI: ZK-based paid remote LLM calls and private Ethereum RPC reads, both of which need inference that doesn't leak the user's prompts or on-chain activity to a third party. He called for more Ethereum-tuned AI models to audit smart-contract and protocol-code security.
Market impact
The 2-bit quantization is the technical beat — it compresses a frontier-class model into a footprint consumer hardware can actually hold, which is the precondition for any meaningful local-inference narrative. Watch the open-source release cadence: if a 90 GB model is the floor today, follow-on quantizations are likely to push that number down quickly, widening the addressable hardware base for private on-chain AI tooling.
Frequently asked questions
-
What did Vitalik Buterin say about DeepSeek V4?
He shared benchmarks for a 2-bit quantized version of DeepSeek V4 — roughly 90 GB of VRAM, about 35 tokens per second on Apple hardware, and 7 tok/s on AMD hardware — framing it as the first concrete test of consumer-grade private inference.
-
What is CROPS AI?
CROPS AI is Vitalik's framework for private, verifiable inference. He argues it should be hardware-agnostic rather than labelled "decentralized AI," and that it overlaps with a CROPS Ethereum access layer covering ZK-based paid remote LLM calls and private RPC reads.
-
Why does 2-bit quantization matter for local AI?
It compresses a frontier-class model into a ~90 GB footprint that consumer hardware can actually hold. That compression is the precondition for running private inference locally instead of routing prompts through third-party data centres.
-
How does this connect to Ethereum privacy?
Vitalik outlined a CROPS Ethereum access layer that supports ZK-based paid remote LLM calls and private Ethereum RPC reads, both of which need inference that does not leak the user's prompts or on-chain activity to a third party.
-
Did Vitalik call for Ethereum-specific AI models?
Yes. He called for more Ethereum-tuned AI models specifically to audit and improve smart-contract and protocol-code security.
CoinTelegraph
WatcherGuru
Crypto News
CryptoSlate
CoinDesk
Crypto Rank News