Gas Killer

Gas Killer · Live Testnet Demo

A language model,
running on Ethereum.

A 35-billion-parameter transformer in pure Solidity — executed off-chain by a Gas Killer operator quorum, settled on Sepolia with one BLS signature.

35B parameters (Qwen3.5 MoE) ~3.6T gas per answer, simulated 384k gas settled on-chain 9,400,000x execution compressed operator BLS quorum
sharded + native fast-executor — one answer split across the operator committee Qwen3.5-35B ≈75 min per answer Qwen3-0.6B ≈4 min per answer measured live · bit-exact committee agreement per segment
"what is ethereum?"prompt #a75e83…agreed bit-exact ⛓ 1 txsettled

one inference → hash-committed segments across the operator committee → bit-exact agreement → one on-chain settlement

01Step 01 · Ask

Prompt → token ids

"what is…?" 38383743394630

Byte-level BPE runs in your browser — the chain sees raw ids.

02Step 02 · Simulate

The quorum runs the model

OP·0 OP·1 OP·2 ✓ bit-exact

Trillions of gas off-chain — every honest operator computes the identical result.

03Step 03 · Settle

One signature on-chain

3.6T gas 384k ⛓ ✓

One BLS signature on Sepolia. Wrong results are slashable.

On-chain history

loading events from Sepolia…

Bisection proof — disputing a billion-gas inference with log₂(n) hashes

Every Gas Killer result is signed by a quorum, and every honest operator computes a bit-identical integer inference. If a signed result is ever wrong, a challenger doesn't re-execute billions of gas on-chain — that's impossible. Instead the execution is checkpointed at every generated token (the commitment chain the contract already maintains), and challenger + defender bisect: compare the midpoint checkpoint, recurse into the half that disagrees, and after ⌈log₂ n⌉ rounds isolate one token's forward pass — small enough to re-execute inside the SP1 slashing guest. The fraud is proven at the first divergent step; the dishonest party is slashed. This lab simulates the whole game.

Execution trace — one segment per generated token

token 0 (prompt committed)
total execution gas (off-chain)
disputed window (segments)
bisection rounds used / max
gas to prove the culprit segment
vs naive full re-execution

Dispute transcript

How this maps to Gas Killer

Checkpoints are free. The consumer's story/chat root is already a running keccak commitment; emitting the per-token intermediate roots during simulation costs the operators nothing extra — they're deterministic integer states.

The pinned environment is part of the claim. Both parties bisect over the same env_commitment — gas limits and the overlay manifest (0x23216c…7ae9 for Qwen). A defender who simulated with different weights diverges at segment 0 and is slashed immediately: availability games can't forge the manifest.

The endgame is tiny. One Qwen token ≈ 28.6B gas — 50,000× smaller than the naive dispute, and exactly the workload the SP1 guest re-executes to produce a fraud proof bound to chainConfigHash.

Honest defenders can't lose. Try the honest mode: every midpoint matches, the challenger's window never isolates a divergence, and the challenge bond is forfeited instead.

The Gas Killer pipeline

Gas Killer architecture
The same pipeline every Gas Killer consumer uses — this demo just makes the “execution” box a 596M-parameter transformer. More at gaskiller.xyz.