EWE-1-slim-large

EWE-1-slim-large

EWE-1 (Ethereum Wallet Embeddings) is a family of causal transformer models designed to compress complex on-chain behavior into rich, forward-looking vector representations, called the EWE-1 model family. Created by sistemalabs and powered by the sequifier library.

This repository hosts the EWE-1-slim-large model weights, with 488 million parameters.

Model Architecture

The model uses a causal attention mechanism to understand temporal patterns in the behavior of a particular Ethereum address. It ingests 31 contextual, behavioral, and temporal features per transaction (including counterparty addresses, contract invocations, and transaction costs) separately. The concatenated feature representations are fed into the transformer, which is trained to predict the next sequence value for all 31 features simultaneously.

• Embedding Dimension: 1536
• Transformer Layers: 24
• Attention Heads: 16
• Look-back Window: 64 transactions

Training Data

EWE-1-slim-large was trained on a massive dataset of Ethereum activity:

• 1.1 Billion Transactions: Encompassing every transaction on the Ethereum chain from 2024 and 2025.
• Dual Perspectives: Recorded on both the sender wallet transaction sequence and the receiver wallet transaction sequence.
• Filtering: Wallets with histories exceeding 100,000 transactions were excluded to prevent extreme outliers from skewing the model.
• Categorical Mapping: High-cardinality fields (like correspondent_address, final_counterparty, related_contract, and method_id) are preserved for the top 1000 most frequent values per month, with remaining values mapped to an "other" token.

Usage & Inference

To run inference or extract embeddings with this model, you must use the official inference implementation, which handles the necessary data preprocessing and relies on the Sequifier engine.

Get started here: EWE-1 Inference Repository

Ethical Guidelines and Responsible Use

The EWE-1 models are released under the highly permissive Apache 2.0 license. Our goal is to democratize access to blockchain-scale Large Transaction Models (LTMs) and remove the friction that typically prevents Web3 builders and academic researchers from utilizing enterprise-grade behavioral modeling.

While the license itself does not legally restrict fields of endeavor, the creators of EWE-1 strongly urge all users to adhere to the following ethical guidelines.

Intended Use Cases EWE-1 was built to enhance the security, transparency, and understanding of decentralized networks. We encourage the deployment of this model for:

• Web3 Security: Building robust, real-time detection systems for phishing, malicious contract interactions, and emerging fraud vectors.
• Network Analysis: Identifying coordinated inauthentic behavior, bot networks, and Sybil attacks to protect decentralized governance and protocol health.
• Academic Advancement: Furthering research in sequence modeling, graph networks, and temporal dynamics within blockchain ecosystems.

Ethical Boundaries Blockchain data is public, but the individuals behind the addresses have a reasonable expectation of safety. As an advanced causal embedding model, EWE-1 is highly capable of separating and clustering wallet behaviors. We expect adopters to respect the following boundaries:

• Respect for Privacy: Embeddings should not be used for unconsented deanonymization, doxxing, or attempting to link pseudonymous on-chain activity to real-world identities for the purpose of harassment or unauthorized surveillance.
• Avoidance of Malicious Targeting: The model must not be utilized to profile vulnerable users in order to craft targeted social engineering campaigns, phishing attacks, or predatory financial exploits.

Model Limitations and Biases To use EWE-1 responsibly, practitioners must account for its inherent limitations:

• Temporal Scope: The model was trained exclusively on Ethereum mainnet transactions from 2024 and 2025. Given the rapid evolution of DeFi protocols and smart contracts, the model’s predictive accuracy and behavioral baselines may degrade when applied to current or future data outside this window.
• Representation Bias: The training dataset explicitly excludes wallets with fewer than 4 transactions or more than 100,000 transactions over their lifetime. As a result, the embeddings may not accurately capture the behavior of brand-new users, extreme high-frequency trading bots, or institutional mega-wallets.
• Not Financial Advice: EWE-1 is purely a behavioral sequence model. It is not designed to forecast market dynamics, predict token valuations, or generate automated trading signals.

By using EWE-1, you join us in the responsibility of advancing Web3 AI safely. We encourage developers to implement human-in-the-loop reviews for any automated system making critical security or moderation decisions based on these embeddings.

Links & Resources

• EWE-1 Project Website: ewe-1.com
• Creators: sistemalabs.com
• Inference Code: github.com/0xideas/ewe-1-inference
• Powered by Sequifier: github.com/0xideas/sequifier