y0sif/Arcwright

Arcwright: Specialized Rust Code Generation Model

Arcwright is a 7.9 billion parameter model, built on the Gemma 4 E4B-it architecture, and uniquely fine-tuned for generating code in modern Rust web and AI frameworks. It focuses on Leptos, Axum, and Rig, making it a highly specialized tool for Rust developers.

Key Capabilities

• Superior Rust Code Generation: Achieves an impressive 6.87 / 10 overall on the RustWebBench-15 benchmark, specifically designed to evaluate Rust web and AI framework code.
• Outperforms Larger Models: Surpasses its 6.5 times larger base model (Gemma 4 26B-A4B) and other prominent models like Claude Haiku and Gemini in its specialized domain.
• High-Quality Training Data: Trained on a meticulously curated dataset (y0sif/Arcwright-v4-Combined) that includes compile-verified code examples for Leptos, Axum, and Rig, ensuring high accuracy and idiomatic output.
• Robust Training Methodology: Utilizes a 3-gate quality pipeline for training data, including sub-agent generation, LLM judging, and cargo check compile verification, to prevent hallucinated APIs.
• Catastrophic Forgetting Prevention: Incorporates a general Rust replay buffer (28% of training mix) to maintain broad Rust knowledge while specializing.

Good For

• Rust Web Development: Ideal for generating components and logic using the Leptos and Axum frameworks.
• Rust AI Framework Development: Useful for working with the Rig AI agent framework, though with some limitations in niche API surface.
• Developers Seeking Idiomatic Rust Code: Provides code that is not only correct but also follows common Rust patterns and best practices.
• Benchmarking and Research: The training data, evaluation prompts, and judge rubric are publicly available via y0sif/OxideCoder for further research and development.