Qwen3.6-35B-A3B-8bit-qpack

Qwen3.6-35B-A3B 8-bit, Swiftlet streaming container (.qpack)

The higher-quality Mac tier of the streamed 35B: 34 GB on disk, about 7.6 GB of RAM, no 4-bit quantization artifacts.

This is Qwen3.6-35B-A3B (8-bit MLX quantization from mlx-community/Qwen3.6-35B-A3B-8bit, 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.

Why 8-bit

Tested head to head on identical prompts, the 8-bit build fixes the repetition loops the 4-bit build shows on longer writing tasks (emails, multi-paragraph text), and stops cleanly where 4-bit runs into its token cap. Factual recall is unchanged (only about 3B parameters are active per token at any precision). If you have the disk and RAM, this is the better Mac experience; the 4-bit container remains the phone-sized tier.

Use it

Swiftlet CLI (macOS):

swiftlet-repack --from-url https://pub-c0cfece2dbc340dbb2cd9d94310a7d68.r2.dev/qwen3.6-35b-8bit-qpack \
  --output qwen3.6-35b-8bit.qpack
swiftlet chat qwen3.6-35b-8bit.qpack "Write a short email asking my landlord to fix a leaking tap."

(The R2 mirror above is usually faster than an anonymous Hugging Face download; --from-hf Leonickson/Qwen3.6-35B-A3B-8bit-qpack works too.)

Status

Early alpha. Measured on an M5 MacBook: about 3.8 tok/s decode, 7.6 GB peak RAM at a 2 GB expert cache. Mac only; it does not fit the iPhone memory budget. Performance work is ongoing.

Siblings: Qwen3.6-35B-A3B-qpack (4-bit, 18 GB disk, 2.6 GB RAM, runs on iPhone) and Qwen3-Next-80B-A3B-qpack (4-bit, 42 GB disk, 4.3 GB RAM).

Note: Hugging Face's parameter widget reads only model.safetensors (the small resident core). 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, tokenizer files, hashes.json (for mirror verification with scripts/verify_container.py).

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.