Qwen3.6-35B-A3B-qpack

Qwen3.6-35B-A3B, Swiftlet streaming container (.qpack)

Run a 35-billion-parameter model on an iPhone or any Apple Silicon Mac in about 2.5 GB of RAM.

This is Qwen3.6-35B-A3B (4-bit MLX quantization from mlx-community/Qwen3.6-35B-A3B-4bit, weights byte-identical, no re-quantization) repacked into the Swiftlet .qpack streaming container: the dense core stays resident while the model's 10,240 routed experts stream from storage on demand, one read per expert, so the 18 GB model runs inside a bounded memory budget.

Use it

• Priv AI on the App Store: open Settings, then Experimental Models, and tap Download. The app streams this repo straight into the on-device container (resumable). The feature ships in the newest app version; if it is still in App Store review you may not see it yet, check back in a couple of days.
• Swiftlet CLI (macOS):

swiftlet-repack --from-hf Leonickson/Qwen3.6-35B-A3B-qpack --output qwen3.6-35b.qpack
swiftlet chat qwen3.6-35b.qpack "Who wrote One Hundred Years of Solitude?"

Status

Early alpha. Measured: 7 to 11 tok/s decode on an M5 MacBook (2.6 GB peak RAM at a 1 GB expert cache), about 1 tok/s on iPhone 17 with a 30 second first reply. Performance work is ongoing. Only about 3B parameters are active per token, so expect strong chat fluency and small-model factual recall.

Bigger sibling for Macs: Qwen3-Next-80B-A3B-qpack (42 GB disk, 4.3 GB RAM, 4.5 to 5 tok/s on M5).

Note: Hugging Face's parameter widget reads only model.safetensors (the small resident core, about 0.4B params). The full 35B lives in the packed_experts/ blobs, see Files.

Layout

model.safetensors (resident dense weights) + packed_experts/layer_XX.bin (fixed-stride expert blobs; one pread per expert) + manifest.json, packed_experts/layout.json, tokenizer files.

Credits and license

Weights: Alibaba Qwen team, Apache 2.0. Quantization: mlx-community. Container format inspired by TurboFieldfare and colibrì. Redistributed under Apache 2.0 with all upstream terms preserved.