XpressAI/Qwen3.5-2B-RYS-UD-Q4_K_XL-GGUF

Qwen3.5-2B — RYS Layer Surgery (GGUF)

Two modified versions of Qwen3.5-2B-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
Qwen3.5-2B-UD-Q4_K_XL.gguf	24	1.34 GiB
Qwen3.5-2B-rys_8-11-UD-Q4_K_XL.gguf	28	1.39 GiB
Qwen3.5-2B-rys_7-10_reasoning-UD-Q4_K_XL.gguf	28	1.38 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 (24 layers)	0.188	0.0	0.118
rys_8-11 (28 layers)	0.125	11.3	0.176
rys_7-10_reasoning (28 layers)	0.062	0.0	0.294

• 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_8-11 shows the best combined improvement: EQ rises from 0 to 11.3 and reasoning improves. rys_7-10_reasoning achieves the highest reasoning score (0.294 vs 0.118 baseline) but at the cost of math and EQ.

---

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 → … → 7 → 8 → 9 → 10 → 11 → 12 → … → 23
rys_8-11: 0 → … → 7 → 8 → 9 → 10 → 11
                      → 8 → 9 → 10 → 11 → 12 → … → 23

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

---

Hybrid Mamba/attention architecture constraint

Qwen3.5-2B is a hybrid SSM/attention model (full_attention_interval = 4): full attention every 4th layer, Gated DeltaNet SSM everywhere else. The architecture repeats 6 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 → 24 + 4 = 28 layers (28 ÷ 4 = 7 ✓)
• Block size 8 → 24 + 8 = 32 layers (32 ÷ 4 = 8 ✓)
• Block size 3 → 24 + 3 = 27 layers (27 ÷ 4 = 6.75 ✗ → crash)

rys_8-11 duplicates layers 8–11, one complete DeltaNet+Attention unit. rys_7-10_reasoning duplicates layers 7–10, spanning the boundary between two attention intervals.

---

How the circuit was found

A two-pass sweep over the 24-layer model:

Pass 1 — 8-layer blocks, stride 4, layers 0–16:

• (4, 12) identified as the hot zone: EQ=11.76, reasoning=0.176

Pass 2 — 4-layer blocks, stride 1, layers 4–16:

• (8, 12) best combined: EQ=11.33, reasoning=0.176
• (7, 11) best reasoning: reasoning=0.294

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 combined (EQ + reasoning)
llama-server -m Qwen3.5-2B-rys_8-11-UD-Q4_K_XL.gguf -ngl 99 --port 8080

# Best reasoning
llama-server -m Qwen3.5-2B-rys_7-10_reasoning-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 ~1.4 GiB (Q4_K_XL, 28 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 Qwen3.5-2B
• llama.cpp for local inference

License

Apache 2.0 (inherited from base model)