Qwen/Qwen3.5-2B, packed for the SiLVANN engine (the Eval-Apply GPU Interpreter and its neural
network library): every layer matrix at 4 bits a weight (TurboQuant, a per-row scale), each rotated along its input axis with a blocked Hadamard-and-signs rotation; the embedding and head at 8 bits — the head made from the embedding, which the model ties it to; norms and other small tensors kept lossless. D4: dense 4 (the model has no experts). Text only: the vision tower and the multi-token prediction layer are left out (pack.json lists them), so that it loads within what a phone leaves free. About 1.7 GB.
The format is silvann-packed-v3 — pack.json describes the codec and the per-tensor width policy, layers/NNN/
holds each layer's records and global.pkl the embedding, head and final norm. It is read by SiLVANN's runtime, not
by transformers. The tokenizer, chat template and configs are the original model's.
Checked against Hugging Face's own forward of the original weights: the greedy continuation token for token, its top five logits 0.35 off on average.
Measured: on a phone — a Razer Edge (Snapdragon G3x Gen 1), Android 12, in Termux — 7.0-7.7 tokens a second with its
weights multiplied in integers (SILVANN_ARM_INT8=1), 4.5 without; on the CPU alone, 7.6 on four threads of a Xeon
E5-2680 v4. On a phone: SiLVANN's docs/android.md (experimental).
All credit for the model belongs to the Qwen team; this repository only changes how its weights are stored. The original license (Apache 2.0) applies and is included.
layers/000/weights.bin 30 MBlayers/001/weights.bin 30 MBlayers/002/weights.bin 30 MBlayers/004/weights.bin 30 MBlayers/005/weights.bin 30 MBlayers/006/weights.bin 30 MBlayers/008/weights.bin 30 MBlayers/009/weights.bin 30 MBlayers/010/weights.bin 30 MBlayers/012/weights.bin 30 MBlayers/013/weights.bin 30 MBlayers/014/weights.bin 30 MBlayers/016/weights.bin 30 MBlayers/017/weights.bin 30 MBlayers/018/weights.bin 30 MBlayers/020/weights.bin 30 MBlayers/021/weights.bin 30 MBlayers/022/weights.bin 30 MBlayers/003/weights.bin 26 MBlayers/007/weights.bin 26 MBlayers/011/weights.bin 26 MBlayers/015/weights.bin 26 MBlayers/019/weights.bin 26 MBlayers/023/weights.bin 26 MBQwen/Qwen3.5-2B