Leonickson/Qwen3-Next-80B-A3B-qpack

Qwen3-Next-80B-A3B-Instruct, Swiftlet .qpack container (4-bit)

This is Qwen3-Next-80B-A3B-Instruct (MLX 4-bit quantization, group size 64) repacked into the .qpack streaming container for Swiftlet, a Swift + Metal runtime that runs large Mixture-of-Experts models on ordinary Apple Silicon by streaming experts from storage instead of loading them into RAM.

Measured on a base M5 MacBook (24 GB):

Metric	Value
Disk	42 GB
Peak RAM	4.3 GB (2 GB expert cache)
Decode speed	4.5 to 5.3 tok/s
First token	about 5 s on a short prompt

This model is for Macs. iPhones do not have the storage for it; the phone model is Qwen3.6-35B-A3B-qpack.

How to use

git clone https://github.com/leonickson1/Swiftlet.git && cd Swiftlet
swift build -c release

# Download this container (resumable):
.build/release/swiftlet-repack \
  --from-hf Leonickson/Qwen3-Next-80B-A3B-qpack \
  --output ~/models/qwen3-next-80b.qpack

# Chat:
.build/release/swiftlet chat ~/models/qwen3-next-80b.qpack \
  "Who are the main characters in The Big Bang Theory?"

Notes

• The parameter count shown in the Hugging Face sidebar reflects only the small resident file (model.safetensors); the routed experts live in packed_experts/ and bring the true total to 80B parameters.
• Weights are byte-identical to the MLX community 4-bit conversion; the container only reorganizes them for single-read expert access.
• Weights are governed by the Qwen license terms (Apache 2.0).