Qwen3.5-9B-RYS-0-4 (GGUF)
An experimental model created by duplicating layers 0-3 in Qwen3.5-9B using the
RYS (Repeat Your Self) method.
What This Is
The base Qwen3.5-9B model (32 layers) with layers 0-3 physically duplicated, creating a 36-layer model. The forward pass executes: layers 0,1,2,3 → 0,1,2,3 (again) → 4,5,...,31.
No training, no fine-tuning, no weight changes — just architectural routing.
Method
1python layer_path.py Qwen3.5-9B-Q4_K_M.gguf \
2 Qwen3.5-9B-RYS-0-4.gguf \
3 -p "0..3,0,1,2,3,4..31" -v
Architecture Discovery
Qwen 3.5 uses a hybrid [DeltaNet, DeltaNet, DeltaNet, Attention] repeating pattern (4-layer cycles). Layer duplication must use block sizes that are multiples of 4 to preserve this pattern — block size 3 crashes with "missing tensor" errors.
This is the first known application of the RYS method to a hybrid DeltaNet/Attention architecture.
Fair Evaluation Results (max_tokens=4096, no think-tag stripping)
| Model | Code Gen | Hallucination Resistance | Reasoning | Overall |
|---|
| Baseline (32 layers) | 80% | 40% | 100% | 73.3% |
| RYS (0,4) (36 layers) | 60% | 80% | 100% | 80.0% |
Tested with 15 questions (5 code generation, 5 hallucination detection, 5 reasoning). Both models evaluated under identical conditions.
Key finding: The improvement is primarily in hallucination resistance (+40 percentage points). Code generation shows a tradeoff (-20 percentage points). Reasoning is unchanged.
Important caveats:
- Sample size is small (5 questions per category) — results need validation with larger benchmarks
- The improvement may be related to response generation behavior rather than capability differences
- The original RYS method was validated on standard transformer architectures (72B+ models); Qwen 3.5's hybrid DeltaNet architecture is untested territory
- Side-by-side testing on harder hallucination prompts showed identical responses from both models
Circuit Map of Qwen 3.5-9B
Full sweep results (7 configurations tested):
| Config | Code | Hallucination | Reasoning | Overall | vs Baseline |
|---|
| Baseline | 80% | 40% | 100% | 73.3% | — |
| (0,4) | 60% | 80% | 100% | 80.0% | +6.67% |
| (4,8) | 80% | 60% | 80% | 73.3% | +0.00% |
| (8,12) | 0% | 40% | 80% | 40.0% | -33.33% |
| (12,16) | 0% | 60% | 80% | 46.7% | -26.67% |
| (16,20) | 0% | 40% | 100% | 46.7% | -26.67% |
| (20,24) | 60% | 60% | 100% | 73.3% | +0.00% |
Usage
1# With llama.cpp
2llama-server -m Qwen3.5-9B-RYS-0-4.gguf -c 8192 -ngl 99
3
4# With Ollama (create a Modelfile)
5echo 'FROM ./Qwen3.5-9B-RYS-0-4.gguf' > Modelfile
6ollama create qwen35-rys -f Modelfile
7ollama run qwen35-rys
Specifications
- Base model: Qwen/Qwen3.5-9B
- Quantization: Q4_K_M
- Original layers: 32
- Modified layers: 36 (layers 0-3 duplicated)
- File size: ~6.2 GB
- Extra VRAM: ~0.5 GB over base model
- Inference overhead: ~12% slower
References
License
Apache 2.0 (same as base model)
