XpressAI/Qwen3.5-0.8B-RYS-UD-Q4_K_XL-GGUF

Qwen3.5-0.8B — RYS Layer Surgery (GGUF)

A modified version of Qwen3.5-0.8B-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
Qwen3.5-0.8B-UD-Q4_K_XL.gguf	24	533 MiB
Qwen3.5-0.8B-rys_4-11-UD-Q4_K_XL.gguf	32	621 MiB

---

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)	0.062	0.0	0.000
rys_4-11 (32 layers)	0.000	36.3	0.353

• 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

Note: the math score drops to 0 in the RYS model. The base model's math score (0.062) is already very low for a 0.8B model.

---

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 → 1 → … → 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10 → 11 → 12 → … → 23
rys_4-11: 0 → 1 → … → 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10 → 11
                           → 4 → 5 → 6 → 7 → 8 → 9 → 10 → 11 → 12 → … → 23

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

---

Hybrid Mamba/attention architecture constraint

Qwen3.5-0.8B 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)

The winning circuit (layers 4–11) spans exactly two complete SSM+Attention units, which appear to form a coherent functional block. No 4-layer sub-block within this range produced comparable improvements in the fine-grained sweep.

---

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) dominant: EQ=36.3, reasoning=0.353 vs baseline EQ=0.0, reasoning=0.0

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

• No 4-layer sub-block matched the 8-layer result
• Best 4-layer blocks: (11,15) reasoning=0.235, (7,11) EQ=4.73

The 8-layer block is the minimum effective circuit size for this model.

---

Usage

llama.cpp / llama-server

llama-server -m Qwen3.5-0.8B-rys_4-11.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

The model weights are ~621 MiB. 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 Qwen3.5-0.8B
• llama.cpp for local inference

License

Apache 2.0 (inherited from base model)