Bytecode Rails LLMs
huggingface.co/bytecodehrPublic Hugging Face models fine-tuned for idiomatic Ruby on Rails code generation.
Problem: General coding models often miss team conventions and generate Rails code that looks plausible but does not match how mature Rails products are actually built and maintained.
What I did:
- Built and published Bytecode’s Rails-focused Hugging Face models
- Fine-tuned Qwen3-based models on 111,000 samples from internal Rails projects
- Prepared GGUF/Ollama-friendly releases for local development workflows
- Documented the dataset engineering, training, quantization, and deployment process
Result: Public Rails-specialized LLMs, including an 8B laptop-friendly model and a 31B MoE flagship model, that encode Bytecode’s Rails conventions directly into the weights.
Tech: Hugging Face, Qwen3, QLoRA, Unsloth, GGUF, Ollama, Ruby on Rails
Rails-tuned models
Run locally with Ollama or use the adapters from Hugging Face.
qwen3-coder-30b-rails
Production-grade Rails code generation. MoE architecture — only 3B params active per token.
View on Hugging Faceqwen3-8b-rails
Lightweight model that runs on a laptop. 5 GB download, 8 GB RAM minimum.
View on Hugging Faceqwen2.5-coder-7b-rails
LoRA adapter for Qwen 2.5 Coder 7B. Merge with base model for Rails-tuned generation.
View on Hugging Faceqwen2.5-coder-3b-rails
Compact LoRA adapter. Fast inference for code completion and small generation tasks.
View on Hugging Face