This model was trained to produce detailed chain-of-thought reasoning inside <think> tags before giving its final answer, distilled from Claude Opus 4.6 and Qwen3.5 reasoning traces.

Quantizations

File	Quant	Size	Description
`qwen3.5-9b-opus4.6-distill-Q2_K.gguf`	Q2_K	~3.5 GB	Smallest, lowest quality. For very constrained devices.
`qwen3.5-9b-opus4.6-distill-Q3_K_M.gguf`	Q3_K_M	~4.5 GB	Low quality, usable for testing.
`qwen3.5-9b-opus4.6-distill-Q4_K_M.gguf`	Q4_K_M	~5.5 GB	Recommended. Best balance of quality and size.
`qwen3.5-9b-opus4.6-distill-Q5_K_M.gguf`	Q5_K_M	~6.5 GB	High quality, moderate size.
`qwen3.5-9b-opus4.6-distill-Q6_K.gguf`	Q6_K	~7.5 GB	Very high quality, near-lossless.
`qwen3.5-9b-opus4.6-distill-Q8_0.gguf`	Q8_0	~9.5 GB	Highest quality quantization.
`qwen3.5-9b-opus4.6-distill-f16.gguf`	F16	~18 GB	Full precision, no quantization loss.

For most users, Q4_K_M or Q5_K_M is the sweet spot.

How to Use

llama.cpp

llama-cli -m qwen3.5-9b-opus4.6-distill-Q5_K_M.gguf -p "<|im_start|>system\nYou are a deep reasoning AI. Think carefully inside <think> tags before answering.<|im_end|>\n<|im_start|>user\nExplain why the sky is blue.<|im_end|>\n<|im_start|>assistant\n" -n 2048

Ollama

ollama run empero-ai/qwen3.5-9b-opus4.6-distill

LM Studio / GPT4All / Jan

Download the GGUF file of your choice and load it directly in the application.

Training Details

Method

Stage 1 — SFT (Supervised Fine-Tuning): 3 epochs on ~13K examples teaching the model the <think> reasoning format using QLoRA (4-bit, rank 64, alpha 128)
Base model: Qwen/Qwen3.5-9B
Hardware: RTX 5090 (32GB VRAM)
Attention: SDPA
Optimizer: Paged AdamW 8-bit
Learning rate: 1e-4 with cosine schedule
Effective batch size: 8 (batch 1 × gradient accumulation 8)
Max sequence length: 4096

SFT Results

Metric	Epoch 1	Epoch 2 (best)	Epoch 3
Eval Loss	0.5205	0.4809	0.4915
Eval Token Accuracy	0.8494	0.8615	0.8617
Eval Entropy	0.508	0.434	0.394

Best checkpoint (epoch 2) was selected via load_best_model_at_end.

Datasets

Dataset	Examples	Type
nohurry/Opus-4.6-Reasoning-3000x-filtered	2,326	Problem → thinking → solution
Jackrong/Qwen3.5-reasoning-700x	633	ShareGPT with `<think>` tags
TeichAI/claude-4.5-opus-high-reasoning-250x	250	Messages with `<think>` tags
Roman1111111/claude-opus-4.6-10000x	9,631	Messages with reasoning traces
Total	12,840

Output Format

The model outputs reasoning in <think> tags followed by its final answer:

<think>
The user is asking about why the sky appears blue. This involves Rayleigh scattering...
 
Sunlight contains all wavelengths of visible light. When it enters Earth's atmosphere,
shorter wavelengths (blue/violet) scatter more than longer wavelengths (red/orange)...
 
While violet actually scatters more than blue, our eyes are more sensitive to blue light,
and some violet is absorbed by the upper atmosphere...
</think>
 
The sky appears blue due to Rayleigh scattering. When sunlight passes through Earth's
atmosphere, the shorter blue wavelengths scatter in all directions more than the longer
red wavelengths. Although violet light scatters even more, our eyes are more sensitive
to blue, and some violet is absorbed higher in the atmosphere — so we perceive the sky
as blue.

About Empero AI

This model was developed by Empero AI. We build open-source AI tools and models focused on advancing reasoning capabilities in smaller, efficient language models.

License

This model inherits the Apache 2.0 license from Qwen3.5-9B.

Dzluck/Qwen3.5-9B-Claude-Opus-4.6-Distill-GGUF