🚀 RELEASES & LAUNCHES
- ExL3-QLoRA enables training 2–8 bpw ExLlamaV3 quantized models
UnstableLlama released ExL3-QLoRA for native QLoRA training of 2–8 bpw ExLlamaV3 models.- The tool targets developers fine-tuning local ExLlamaV3 models.
- ExL3-QLoRA uses plain PyTorch without Transformers and supports 2–8 bpw quantized models.
- It supports Llama 3, Mistral, Qwen3, Qwen3.5–3.8, Gemma 3–4, and Arcee Trinity families.
- Adapters save in standard PEFT format, load in TabbyAPI or PEFT, and can be merged into the base model.
- The project expands QLoRA beyond the previous 4-bit bitsandbytes constraint for ExLlamaV3 users.