Corpus.CoreLocal LLM for Ethereum
Local LLM for Ethereum
A local LLM for Ethereum is a language model that runs on the user's device — in the browser or in an app — and explains what a transaction will do after it has been simulated on cryptographically verified state. The prompt stays on the machine. Corpus Core's Transaction Security Agent (TSA) is a working prototype of that local SLM for Web3.
General chatbots can talk about crypto from training data. A local SLM for Web3 explains a decoded simulation it was given: events, calls, and state changes, shown next to facts you can check. The model does not decide whether the transaction is safe.
What is a local SLM for Web3?
A local SLM (small language model) for Web3 runs in the browser or app and explains an Ethereum transaction after Colibri has simulated it on verified state. Nothing in the explanation step has to leave the device for a cloud model.
TSA's default local model is colibri-tsa-4b, a fine-tune of Qwen3.5-4B (Apache 2.0, February 2026), trained on decoded simulation results and matching Solidity. WebLLM runs the quantized model on the GPU through WebGPU in current Chrome or Edge. The first download is about 2.4 GB and is cached; the packaged 16k context needs about 3.9 GB of VRAM.
Simulate on verified state, then explain
Before any model runs, Colibri simulates the transaction with colibri_simulateTransaction. It fetches Merkle proofs for touched accounts and storage, verifies them against a consensus-proven state root, and executes locally in its own EVM.
Verified Solidity from Sourcify is used only when compiled bytecode matches the proven codeHash. The prompt then contains decoded events, calls, state changes, and the functions that actually ran — not a guess from raw calldata alone.
- Simulation on proven state — not blind RPC eth_call
- Decode with ABI and storage layout when source matches
- Explain only executed code paths
- Decoded facts stay visible beside the model text
Local LLM vs cloud LLM for wallets
The same explainer pipeline can call OpenAI-compatible APIs, Anthropic, or a local Ollama server. Cloud paths send the decoded prompt off-device. The local WebLLM path keeps explanation on the GPU in the browser.
Local inference is not the same as fully offline. With chainId set, enrichment still fetches public Sourcify metadata. Omit chainId when the explanation must not touch the network. Proof fetches for simulation still contact an RPC or prover for the accounts and slots touched — PAP addresses that remaining read leak.
When a local Ethereum LLM helps
- Wallet and dApp flows that should explain intent before signing
- Teams that want simulation-backed explanations without shipping prompts to a vendor by default
- Browsers with WebGPU that can cache a ~2.4 GB model
- Apps that already verify state with Colibri or vRPC
Local SLM for Web3 vs general local LLM
| General local LLM | Local SLM for Web3 (TSA) | |
|---|---|---|
| Input | User chat / training knowledge | Decoded simulation on verified state |
| Job | Talk about crypto topics | Explain this transaction's proven execution |
| Safety decision | May invent advice | Does not decide if the tx is safe |
| Default runtime | Ollama, desktop apps, etc. | WebLLM in the browser (also Ollama / APIs) |
FAQ
What is a local LLM for Ethereum?
A language model that runs on the user's device and explains an Ethereum transaction after simulation on verified state. TSA is a prototype: a Qwen3.5-4B fine-tune via WebLLM, about 2.4 GB on first download.
What is a local SLM for Web3?
A small language model for Web3 that runs in the browser or app and explains a decoded, proven simulation. The prompt is not sent to a cloud model for that path. See /tsa and the Don't Sign Blind article.
How is this different from a general local LLM?
A general local LLM can discuss crypto from training data. TSA explains a decoded simulation it was given — events, calls, and state changes — and does not decide whether the transaction is safe.
Does the model replace verification?
No. Colibri verifies proofs and simulates first. The model only explains that proven result. Users should still read the decoded facts next to the text.
Where can I try it?
Open the playground at playground.colibri-proof.tech, read the product page at /tsa, or start from the Don't Sign Blind article.
Don't sign blind
Run the TSA pipeline in the browser, or read how a local SLM for Web3 explains a verified simulation.