Qwen3.6-35B-A3B-DBF-SIGN1

Qwen3.6-35B-A3B — DBF/SIGN1 Routed Experts

Post-training structural binarization of all 30,720 routed FFN expert matrices in Qwen3.6-35B-A3B. Each routed projection is represented as:

D_out · U_SIGN1 · D_mid · V_SIGN1 · D_in

The binary factors use the custom one-bit SIGN1 GGUF tensor type. All three diagonal factors are FP16. Every non-routed-expert tensor is copied from the same Q2_K_XL common base used by the matched scalar controls.

Important runtime requirement

Stock llama.cpp cannot load this model. Build the exact SIGN1/DBDBD implementation from:

https://github.com/eodus/sign1-llama.cpp

The runtime needs both the custom SIGN1 tensor type and the Qwen routed D_in -> V -> D_mid -> U -> D_out graph.

Storage

At k=1024, the routed DBF payload costs 2.5546875 bits per original expert weight, slightly below Q2_K's 2.625 bpe.

Quality

All models below share the same non-expert remainder and differ only in routed expert tensors.

DBF removes:

• 58.8% of Q2_K's additive PPL damage;
• 57.5% of the Q2_K-to-FP16 gap on the mean-NLL scale;
• 83.1% of the Q2_K-to-Q3_K improvement on the NLL scale.

Local interpolation between Q2_K (2.625 bpe) and Q3_K (3.4375 bpe) gives approximately 3.30-bpe scalar-equivalent quality at an actual 2.555-bpe DBF payload. This interpolation is specific to this model, common remainder, and WikiText-2 evaluation.

Paired full-PPL result: DBF wins 526/580 chunks; Q2_K wins 54; exact two-sided sign-test p = 3.21e-98.

Matrix approximation

Across all 30,720 routed matrices:

Spectral norms use fixed-seed 32-iteration power iteration rather than exact SVD.

Vulkan performance

Radeon 8060S / Ryzen AI Max+ 395, drift-balanced Q2_K → DBF → DBF → Q2_K order, 10 warmups and 20 measured repetitions per cell:

Method provenance

The fixed-dimension DBF factorization core follows the public Boža–Macko implementation, run for 80 iterations at k=1024. The deployed pipeline then performs one final weighted least-squares reprojection of D_mid for the fixed binary signs and outer diagonals before FP16 simulation and export.

• Paper: Vladimír Boža and Vladimír Macko, Addition Is Almost All You Need, TMLR 2026.
• Public reference code: https://github.com/usamec/double_binary
• Reproduction and analysis scripts: https://github.com/eodus/dbf-sign1
• Runtime: https://github.com/eodus/sign1-llama.cpp

Authors and AI collaboration disclosure

Developed by Sasha Shlemov with Drinkins, his calibrated personal AI research and engineering assistant. Drinkins contributed implementation, experiment design, debugging, analysis, source research, and editing. Sasha directed the research, made the final technical and publication decisions, and takes responsibility for the claims.

Status

Private release candidate. Exact commit hashes, GGUF SHA-256, and clean-checkout reproduction commands will be frozen before public release.