1/ This was really cool. For the first time, an LLM ran entirely on the blockchain. Weights, activations, attention, tokens all loaded and calculated onchain in a general purpose ledger. Every step in the process was a standard SVM transaction.

2/ This wasn't just some toy model either. It was Qwen3.5-4B. That model has roughly GPT-o1 or DeepSeek R1 levels of intelligence.

3/ The whole thing ran on @fogo chain (testnet, but main net shortly). It was a 30 token prompt and 52 token completion. Each decode token took approximately 300 seconds at around 30% chain utilization. The run cost the equivalent of about $1 in gas fees

4/ Besides optimizing the hell out of the kernels, it only works with the SVM. Generally this shape of packed compute would be infeasible on an EVM chain. It required a chain with extremely high throughput, cheap gas and low latency. So for right now "only possible on Fogo"

5/ Now generally, even with the world's most performant chain, I wouldn't recommend anyone use a blockchain to run an LLMs. This was a long latency for a small completion But the cool part is once you run the entire LLM on the chain, that means you can run any *part* of the LLM

6/ So, this was a bit of a stunt, but it represents the components that eventually enable trustless crytpto secured intelligence In that world, AI could be consumed trustlessly like any other onchain primitive or protocol. Imagine AI based smart contracts or onchain markets.

7/ Imagine censorship resistant intelligence, running on permissionless markets, transacting in crypto native assets You could even have a future where agents exercise sovereign control over onchain assets through smart wallets, directly paying for their own compute.

8/ Right now it's a fun research/infra project building on top of @fogo. You can follow at @BaranosAI.

Cool stuff, and putting the whole LLM stack on a blockchain is wild.

Thats really amazing, running the entire LLM flow onchain instead of relying on an offchain service really shows how far Fogo is pushing SVM execution. Onchain AI really feel a lot more close to reality.