Rebiha

Rebiha is a managed fine-tuning platform for open-weight language models. Users select a base model from a catalogue that includes Qwen, Gemma, Llama, Mistral, Phi and DeepSeek, choose training parameters, and Rebiha provisions a GPU, runs the QLoRA training job, and returns the finished model.

Output includes a ready-to-run GGUF file for Ollama, llama.cpp or LM Studio, plus the LoRA adapter weights, tokenizer and configuration files for users who want to merge or continue training themselves.

The platform also includes a library of 35 instruction-tuning datasets across domains such as customer support, legal information, real estate, healthcare administration, software development and e-commerce, so users can fine-tune without building training data first.

Base model weights are never modified; training produces a separate adapter.