Qwen3.5-27B-RYS-UD-Q4_K_XL-GGUF

Qwen3.5-27B — RYS Layer Surgery (GGUF)

Two modified versions of Qwen3.5-27B 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

---

Probe scores

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

• 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
• Logic: fraction correct on logical deduction probes only

rys_30-33 shows the best combined improvement across reasoning categories. rys_34-37 shows the highest EQ score but no reasoning improvement over baseline.

---

Benchmarks (based on BFCLv4)

Non-Live Tests

Live Tests

Multi-Turn Tests

Memory Tests (Agentic)

RYS vs Baseline Comparison (All Tests)

---

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 → … → 29 → 30 → 31 → 32 → 33 → 34 → … → 63
rys_30-33:  0 → 1 → … → 29 → 30 → 31 → 32 → 33 → 30 → 31 → 32 → 33 → 34 → … → 63

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

---

Hybrid Mamba/attention architecture constraint

Qwen3.5-27B is a hybrid SSM/attention model (full_attention_interval = 4): full attention every 4th layer, Mamba SSM everywhere else.

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

• Block size 4 → 64 + 4 = 68 layers (68 ÷ 4 = 17 ✓)
• Block size 3 → 64 + 3 = 67 layers (67 ÷ 4 = 16.75 ✗ → server crash at load)
• Block size 8 → 64 + 8 = 72 layers (72 ÷ 4 = 18 ✓)

Only multiples of 4 work as block sizes for this model family.

---

How the circuit was found

A two-pass sweep over the 64-layer model using a probe benchmark:

Pass 1 — 8-layer blocks, stride 4, layers 4–60:

• Identified hot zones at layers 8–16 (reasoning) and 28–40 (EQ/math)

Pass 2 — 4-layer blocks, stride 1, within each hot zone:

• (30, 34) achieved the best combined score: reasoning=0.353, EQ=29.5, logic=1.0
• (34, 38) achieved the highest EQ score: EQ=39.4

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

llama-server -m Qwen3.5-27B-rys_30-33.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 alone are ~21 GiB (Q4_K_XL quantization, 68 layers). A single A100 80GB or H100 runs this comfortably. Consumer GPU setups depend on your llama.cpp version's tensor split support.

---

Credits

• David Ng for the original RYS method
• Unsloth for the base Q4_K_XL GGUF quantization
• Qwen team for Qwen3.5-27B
• llama.cpp for local inference

License

Apache 2.0 (inherited from base model)