Ornstein3.6-35B-A3B-RYS

A RYS-enhanced version of DJLougen/Ornstein3.6-35B-A3B, with an optimal layer duplication applied using the RYS Brain Scanner method from Ng (2026). No weights were modified — a single critical layer was identified and duplicated, yielding a +49% improvement on combined reasoning and instruction-following benchmarks.

See also: DJLougen/Ornstein3.6-35B-A3B (base fine-tune) | DJLougen/Ornstein3.6-35B-A3B-SABER (SABER ablation applied on top)

Support This Work

I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.

Support on Ko-fi

RYS Brain Scan Results

The RYS method (Ng, 2026) performs an exhaustive sweep of all possible layer duplication configurations (i, j), where layers i through j-1 are repeated in the forward pass. Two orthogonal probe sets — hard arithmetic and constrained instruction following — score each configuration. The resulting heatmap reveals the functional circuit boundaries of the transformer's reasoning anatomy.

Optimal Configuration

Metric	Baseline	After RYS (i=10, j=11)	Delta
Math (hard arithmetic)	0.404	0.968	+139%
IFO (instruction following)	0.875	0.938	+7.2%
Combined	1.279	1.905	+49%

Surgery applied: Layer 10 duplicated (single layer)
New layer count: 41 (up from 40)
Sweep: 820 (i, j) configurations evaluated across 40 layers
No weights modified — the duplicated layer is an exact copy

What This Means

Layer 10 sits at 25% depth in the network — right in the region where early feature extraction transitions into abstract reasoning. Duplicating this single layer gives the model a second pass through a critical reasoning circuit, dramatically boosting arithmetic accuracy while preserving instruction-following ability.

Details

Developed by: DJLougen
Architecture: Qwen3_5MoeForCausalLM — Qwen 3.6 MoE with linear + full attention interleaved (Gated Delta Net)
Parameters: 34.66B total, ~3B active (256 experts, 8 active per token)
Hidden size / layers: 2048 / 41 (40 original + 1 duplicated)
Context length: 262,144 tokens
License: Apache 2.0
Base model: DJLougen/Ornstein3.6-35B-A3B
Method: RYS Brain Scanner (Ng, 2026)

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "DJLougen/Ornstein3.6-35B-A3B-RYS"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Explain mixture-of-experts routing in one paragraph."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Citation

If you use this model or the RYS method:

@article{ng2026rys,
  title   = {LLM Neuroanatomy: How I Topped the LLM Leaderboard Without Changing a Single Weight},
  author  = {Ng, David Noel},
  year    = {2026},
  month   = {March},
  url     = {https://dnhkng.github.io/posts/rys/}
}