BITSANDBYTES / MODEL
Four-bit Qwen weights for Transformers and QLoRA fine-tuning.
THE PROJECT, IN PRACTICE
Quantizes language-model linear layers while preserving the vision tower, embeddings, and output head in BF16. Includes the MTP weights; the model card documents measured tradeoffs and runtime limits. This is a quantization of Qwen’s model, not an original base model.
Model formats are not interchangeable. Match the published artifact to the runtime, architecture support, and available memory documented in its card.
The model card identifies the base model, derivation, and license terms. A quantized build remains a derivative of its upstream model.
FROM IDEA TO USE
Read the model or dataset card and its provenance.
Check the runtime or schema and your intended use.
Open the repository and follow its current usage instructions.
greghavens/Qwen3.8-27B-bnb-4bitUse the source page for current files, requirements, and usage examples.
TAKE IT FOR A SPIN
Quantizes language-model linear layers while preserving the vision tower, embeddings, and output head in BF16. Includes the MTP weights; the model card documents measured tradeoffs and runtime limits. This is a quantization of Qwen’s model, not an original base model.
Published artifact. Review its model card for runtime support and inherited license terms.