TLDR
- Vitalik said DeepSeek V4 has a 2-bit quantized version running within about 90 GB VRAM.
- The model reached about 35 tokens per second on Apple hardware, according to Vitalik.
- Vitalik said AMD hardware reached about 7 tokens per second during local AI tests.
- He linked CROPS AI with private Ethereum RPC reads and ZK-based remote LLM calls.
- CryptoQuant noted rising ETH failed transactions and higher exchange inflows.
Vitalik Buterin said DeepSeek V4 local AI can support Ethereum privacy tools, as local models improve across hardware. He cited a 2-bit quantized version that runs within about 90 GB of VRAM. The update also linked CROPS AI with private Ethereum RPC reads, paid remote LLM calls, and safer smart contract development.
DeepSeek V4 Runs Locally As Vitalik Tracks Hardware Gaps
Ethereum co-founder Vitalik Buterin shared new progress on local AI testing. He said DeepSeek V4 now has a 2-bit quantized version available through a GGUF build. The model can run within about 90 GB of VRAM, based on his update.
Buterin said the model works, but speed differs across hardware. He said Apple hardware reached around 35 tokens per second. He also said AMD hardware reached about 7 tokens per second. He framed this gap as a core challenge for wider local AI use. He said true “CROPS AI” should support more than one hardware maker.
He contrasted that with systems described only as “decentralized AI.” Buterin wrote that proper hardware support shows the difference between the two ideas. His update also mentioned other local AI projects. These included messaging-daemon with alpha Telegram support, Lucebox Hub for dense models, and VoxTerm for local AI recording. He said more projects are coming.
CROPS AI Meets Ethereum Privacy Access Layer
Buterin also connected local AI work with Ethereum privacy tools. He said there is overlap between a “CROPS Ethereum access layer” and CROPS AI. One example involved ZK-based paid calls to remote large language models. He said the same method could help private Ethereum RPC reads.
RPC reads allow users and apps to request blockchain data. Private reads may reduce the data exposed during Ethereum use. The post pointed to a broader privacy path for Ethereum users. Remote LLM calls could be paid and verified with zero-knowledge methods.
At the same time, Ethereum access could become more private through similar tools. Buterin also called for Ethereum-tuned AI models. He said application-specific models can help improve secure code work. He cited Leanstral as an example of a smaller model that performs well on Lean code.
ETH Market Data Shows Higher Network Stress
CryptoQuant data added a market backdrop to the AI and Ethereum discussion. Its chart tracked Ethereum price action, exchange inflows, and failed transactions. The chart showed ETH trading below prior cycle highs. The report said rising exchange inflows can show that large holders are moving ETH to exchanges.
This does not prove selling has started. Yet it can place more liquidity where it can be sold quickly. The lower section of the chart showed a sharp rise in failed Ethereum transactions. The 30-day moving average also moved higher. Failed transactions can come from congestion, bots, DeFi errors, or gas issues. CryptoQuant said the mix of failed transactions and exchange inflows may point to caution.
It stated, “This combination of network friction and potential exchange-bound liquidity could possibly indicate a somewhat bearish outlook for the asset in the near term.” For ETH, price action still needs support and follow-through. A move above local highs would confirm stronger momentum. Lower failed transactions and reduced exchange inflows would make the setup cleaner.
Get 3 Free Stock Ebooks
Discover top-performing stocks in AI, Crypto, and Technology with expert analysis.
- Top 10 AI Stocks - Leading AI companies
- Top 10 Crypto Stocks - Blockchain leaders
- Top 10 Tech Stocks - Tech giants
