Ornstein-122-A10B-GGUF

GGUF quantizations of DJLougen/Ornstein-122-A10B — a reasoning-focused fine-tune of Qwen 3.5 122B-A10B (MoE, ~10B active per token) trained on high-quality reasoning traces curated through a custom Drift Diffusion Modeling (DDM) pipeline.

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

---

---

What Makes Ornstein Different

Unlike typical reasoning fine-tunes that use large volumes of synthetic data, Ornstein implements quality-over-quantity:

• Detects degenerate reasoning: Identifies "fake" reasoning that mimics thought without substance (hedging, restating, circling)
• Premium vs. Degenerate split: DDM pipeline cleanly separates premium from degenerate reasoning traces
• High-fidelity curation: Near-perfect AUC separating premium from degenerate reasoning with >99% sensitivity
• MoE efficiency: 122B total parameters with only ~10B active per token — big model reasoning at a fraction of the compute

The model uses <think>...</think> blocks for extended multi-phase reasoning with self-correction and verification before providing final answers.

---

Available Quantizations

Note: Uploads are in progress — more quantizations may be added.

Quantization	Size	Use Case
F16	~261 GB (split)	Full precision, no quality loss
Q4_K_M	~74 GB (split)	Best quality/size trade-off, recommended
Q8_0	~83 GB (split)	Higher precision, minimal quality loss

---

Quick Start

llama.cpp

# Download a quantization (example: F16 split files)
huggingface-cli download DJLougen/Ornstein-122-A10B-gguf --local-dir .

# Run with llama.cpp
./llama-cli -m Ornstein-122-A10B-F16-00001-of-00006.gguf \
  -p "You are a helpful reasoning assistant." \
  --temp 0.6 -n 8192

Ollama

# Create a Modelfile
cat <<EOF > Modelfile
FROM ./Ornstein-122-A10B-F16-00001-of-00006.gguf
PARAMETER temperature 0.6
PARAMETER num_predict 8192
SYSTEM "You are a helpful reasoning assistant."
EOF

ollama create ornstein-122 -f Modelfile
ollama run ornstein-122

LM Studio

1. Download the desired quantization from the Files tab
2. Load it in LM Studio
3. Set context length to 8192 for full reasoning depth

---

Recommended Settings

Parameter	Suggested Value
Temperature	0.6
Top-P	0.95
Max Tokens	8192
Repeat Penalty	1.1

---

Training Details

Parameter	Value
Base Model	unsloth/Qwen3.5-122B-A10B
Architecture	Mixture-of-Experts (122B total, ~10B active)
Method	LoRA (rank 32, alpha 32)
Dropout	0.0
Epochs	1
Learning Rate	1e-4 (cosine schedule, 10% warmup)
Max Sequence Length	8192
Micro Batch Size	1
Gradient Accumulation	4 steps
Weight Decay	0.01
LoRA Targets	q_proj, k_proj, v_proj, o_proj
Framework	Unsloth

---

Intended Use

Designed for tasks requiring structured, multi-step reasoning:

• Mathematics
• Logic problems
• Code analysis
• Scientific problems
• Complex question answering

The MoE architecture makes it practical to run 122B-class reasoning on hardware that couldn't handle a dense model of the same size.

---

Limitations

• Single epoch training means the model retains most base Qwen 3.5 122B-A10B behavior; the fine-tune primarily shapes reasoning style rather than injecting new knowledge
• Language scope: DDM pipeline optimized for English; other languages reflect base model performance
• Edge cases: Extended thinking can occasionally loop on adversarial or highly ambiguous prompts
• Size: Even quantized, the 122B MoE model requires substantial storage and memory

---

Citation

@misc{ornstein122a10b,
  author = {DJLougen},
  title = {Ornstein-122-A10B: DDM-Curated Reasoning Fine-Tune of Qwen 3.5 122B-A10B},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/DJLougen/Ornstein-122-A10B}
}

---

Links

• Full-precision model: DJLougen/Ornstein-122-A10B
• NVFP4 quantization: DJLougen/Ornstein-122-A10B-nvfp4
• Ornstein-27B series: DJLougen/Ornstein-27B | v2
• Base model: Qwen/Qwen3.5-122B-A10B | unsloth/Qwen3.5-122B-A10B