XpressAI/Qwen3.5-4B-RYS-UD-Q4_K_XL-GGUF

Qwen3.5-4B — RYS Layer Surgery (GGUF)

Two modified versions of Qwen3.5-4B-Instruct produced by RYS layer duplication — no training, no weight changes, just routing hidden states through a specific circuit twice.

Based on David Ng's RYS method.

---

Files

File	Layers	Size
Qwen3.5-4B-UD-Q4_K_XL.gguf	32	2.8 GiB
Qwen3.5-4B-rys_27-30-UD-Q4_K_XL.gguf	36	3.0 GiB
Qwen3.5-4B-rys_28-31_eq-UD-Q4_K_XL.gguf	36	3.0 GiB

---

Probe scores

Scores from an internal sweep benchmark run during circuit search. Sample sizes are small — treat these as directional indicators, not definitive benchmarks.

Model	Math	EQ	Reasoning
Base (32 layers)	0.375	0.0	0.529
rys_27-30 (36 layers)	0.188	0.0	0.588
rys_28-31_eq (36 layers)	0.000	80.7	0.000

• Math: Ng's partial-credit scoring on a small GSM8K sample
• EQ: EQ-Bench-style emotional intelligence score (0–100)
• Reasoning: fraction correct across causal, date, logic, navigation, and GSM8K probes

rys_27-30 improves reasoning above the already-strong baseline (0.588 vs 0.529), with GSM8K rising from 0.2 to 0.8. rys_28-31_eq achieves EQ=80.7, the highest EQ score observed across all model sizes in this sweep, but at the cost of reasoning.

---

What is RYS?

Transformers self-organise during training into functional circuits — contiguous blocks of layers that act together. The RYS technique duplicates a specific block in the forward pass using the same weights, with no extra copies on disk beyond the GGUF file overhead:

Normal:     0 → … → 26 → 27 → 28 → 29 → 30 → 31
rys_27-30:  0 → … → 26 → 27 → 28 → 29 → 30
                          → 27 → 28 → 29 → 30 → 31

The model processes the circuit twice, without any weight changes or fine-tuning.

---

Hybrid Mamba/attention architecture constraint

Qwen3.5-4B is a hybrid SSM/attention model (full_attention_interval = 4): full attention every 4th layer, Gated DeltaNet SSM everywhere else. The architecture repeats 8 times:

3 × (DeltaNet → FFN) → 1 × (Attention → FFN)

This creates a hard constraint on layer surgery: the total layer count must remain divisible by 4.

• Block size 4 → 32 + 4 = 36 layers (36 ÷ 4 = 9 ✓)
• Block size 8 → 32 + 8 = 40 layers (40 ÷ 4 = 10 ✓)
• Block size 3 → 32 + 3 = 35 layers (35 ÷ 4 = 8.75 ✗ → crash)

---

How the circuit was found

A two-pass sweep over the 32-layer model:

Pass 1 — 8-layer blocks, stride 4, layers 0–32:

• (12, 20) best combined: EQ=58.4, reasoning=0.412
• (24, 32) best EQ: EQ=63.1

Pass 2 — 4-layer blocks, stride 1, layers 8–32:

• (28, 32) best EQ: EQ=80.7 (highest across all model sizes in this sweep)
• (27, 31) best reasoning: reasoning=0.588 (exceeds baseline of 0.529), GSM8K=0.80

Each configuration was tested by patching the GGUF layer path, loading with llama-server, and scoring with the probe suite.

---

Usage

llama.cpp / llama-server

# Best reasoning
llama-server -m Qwen3.5-4B-rys_27-30-UD-Q4_K_XL.gguf -ngl 99 --port 8080

# Best EQ
llama-server -m Qwen3.5-4B-rys_28-31_eq-UD-Q4_K_XL.gguf -ngl 99 --port 8080

Thinking mode

Qwen3.5 defaults to thinking mode (<think>…</think>). Add /no_think to the system prompt for fast, direct answers:

messages = [
    {"role": "system", "content": "/no_think"},
    {"role": "user",   "content": "Your question here"}
]

VRAM requirements

Model weights are ~3.0 GiB (Q4_K_XL, 36 layers). Runs on any modern GPU or CPU.

---

Credits

• David Ng for the original RYS method
• Unsloth for the base Q4_K_XL GGUF quantization
• Qwen team for Qwen3.5-4B
• llama.cpp for local inference

License

Apache 2.0 (inherited from base model)