This model is packed for NunSpark compatibility
NunSpark targets a simple, unglamorous problem: your Mac has 16 GB of unified memory and you want to run a model that needs 40+ GB of weights. Instead of refusing to load, NunSpark packs a model into one file per transformer layer ("piece"), and streams each piece from disk into a small resident budget just before it's needed, discarding it right after. Combined with disk-aware speculative decoding, this makes large (13–30B, 4-bit) models usable — a few tokens per second instead of not-at-all — on machines that could never hold the model in RAM.
Read more / Try it on : https://github.com/sharma-open-source/Nunspark
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for nunspark/Qwen3-30b-packed
Base model
Qwen/Qwen3-30B-A3B-Base Finetuned
Qwen/Qwen3-30B-A3B Quantized
mlx-community/Qwen3-30B-A3B-4bit