kev-d3-paper is an open-weight model from Lo, released under Creative Commons Attribution 4.0. Its published files total 1.3 MB.
Working draft v6 (2026-10-07), experimental technical report, not peer reviewed. The roughly twofold joint-KL ratio in §5.7 is a provisional result while a matched-denominator recheck finishes; other numbers may still change.
Model Card
By Lo, published under cc-by-4.0, revision 97c06432cda0.
Working draft v6 (2026-10-07), experimental technical report, not peer reviewed. The roughly twofold joint-KL ratio in §5.7 is a provisional result while a matched-denominator recheck finishes; other numbers may still change. kev-d3-paper-draft-v6.pdf (current) Per-module sub-3-bit sensitivity of two Qwen3.5 hybrid decision models (Gated DeltaNet linear attention interleaved 3:1 with full attention), measured as option-distribution KL under binary / ternary / 3-bit / 4-bit GPTQ, with a multiple-choice knapsack priced in the target runtime's storage format (HyBitQ) that turns the measurements into bit allocations. Related release: Jakevin/kev-4b-ternary-mlx (revision v2.0 uses this…
Read Lo's full model card
Where Can Qwen3.5 Decision Models Afford Low Precision? A Sub-3-Bit Sensitivity and Runtime-Priced Allocation Study (working draft)
Working draft v6 (2026-10-07), experimental technical report, not peer reviewed. The roughly twofold joint-KL ratio in §5.7 is a provisional result while a matched-denominator recheck finishes; other numbers may still change.
kev-d3-paper-draft-v6.pdf (current)
Per-module sub-3-bit sensitivity of two Qwen3.5 hybrid decision models (Gated DeltaNet linear attention interleaved 3:1 with full attention), measured as option-distribution KL under binary / ternary / 3-bit / 4-bit GPTQ, with a multiple-choice knapsack priced in the target runtime's storage format (HyBitQ) that turns the measurements into bit allocations.
Related release: Jakevin/kev-4b-ternary-mlx (revision v2.0 uses this allocation; v1.0 is the earlier hand-designed recipe).
| Kev-4B | decision-v7 retention | held-out retention | size |
|---|---|---|---|
| v1.0 (hand recipe + rank-16 adapter) | 96.8% | 86.5% | 1.679 GB |
| v2.0 (measured allocation, no adapter) | 99.2% | 94.5% | 1.678 GB |
Earlier versions
- kev-d3-paper-draft-v2.pdf: superseded draft v2 (2026-10-07, earlier title "Where Can a Hybrid Linear-Attention Model Afford One Bit?", still contained [TBD] sections). Kept so existing links keep working.
Identity and Version
- Repository
- Jakevin/kev-d3-paper
- Publisher
- Lo
- Task
- Not stated by the source
- Modality
- Other
- Library
- mlx
- Parameters
- Not stated by the source
- Languages
- mlx
- Revision
- 97c06432cda0819a2bdae5140d537d30618c57de
- First published
- 2026-10-07
- Last updated
- 2026-10-07
Files and Weights
4 files, 1.3 MB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| README.md | Documentation | 1.6 KB | — |
| kev-d3-paper-draft-v2.pdf | Other | 579.4 KB | 848e25170a6d |
| kev-d3-paper-draft-v6.pdf | Other | 708.3 KB | 5c4b1781aff4 |
| .gitattributes | Repository | 1.6 KB | — |
License and Download
- License
- cc-by-4.0
- Access
- Open weights, no gate
Released by Lo through its official repository on Hugging Face. Read the license.
Questions About kev-d3-paper
Can I use kev-d3-paper commercially?
Yes. kev-d3-paper is released under Creative Commons Attribution 4.0. CC BY 4.0 permits sharing and adapting the work, including commercially, provided the creator is credited and changes are indicated.