Original · 2059739380785271033 · vitalik.eth @VitalikButerin

Updates since then:

* Deepseek v4 is out. There *is* a 2-bit quant that can run within 90 GB (
https://
huggingface.co/antirez/deepse
ek-v4-gguf
… ), and it works, however it's only fast on Apple hardware (I've head ~35 tok/s). On AMD, it's ~7 tok/s. IMO actually taking the effort to properly support more than one hardware manufacturer is a great example of the difference between mere "decentralized AI" and genuine "CROPS AI". I hope we can become better at this.
*
https://
github.com/vbuterin/messa
ging-daemon
… also has alpha telegram support now. However, the path to adding your account is quite janky
*
https://
github.com/Luce-Org/luceb
ox-hub
… looks promising as a way to run "dense" models (eg. Qwen 27B) more efficiently. It's janky, but on my 5090 laptop it seems to be ~2x more tok/s than llama.cpp
* VoxTerm (local AI recording, no third-party servers) continues to be developed
https://
github.com/dmarzzz/VoxTerm

And there's a lot more projects coming on the horizon.

One other thing that has been on my mind is that there's actually a lot of intersection between "CROPS ethereum access layer" and "CROPS AI". For example, we want a ZK way to make (paid) calls to remote LLMs. But if we have this, then it's just as useful for solving another problem: private RPC reads in Ethereum.

Another example: application-specific finetuned LLMs. Leanstral (
https://
mistral.ai/news/leanstral ; I get ~38 tok/s on AMD) fits into < 70 GB, but can hold its own against 1T models on writing Lean code. Things like this are a huge boon for writing more secure code (
https://
vitalik.eth.limo/general/2026/0
5/18/fv.html
… ). We should have models finetuned for Ethereum-related use cases as well.

RT @VitalikButerin: Updates since then:

* Deepseek v4 is out. There *is* a 2-bit quant that can run within 90 GB ( https://t.co/X3AFAsiH02…