rico03/Qwopus-GLM-9B-DualReason-Distilled-GGUF

Qwopus-GLM-9B-DualReason-Distilled

A DARE-TIES weight merge of two specialized Qwen3.5-9B fine-tunes, combining Opus-style agentic reasoning with GLM-5.1 structured chain-of-thought into a single 9B model.

This model is the result of an extensive research process that explored ANN-based layer routing before converging on DARE-TIES as the optimal merging strategy. The full methodology and findings are documented below.

---

Benchmark Results

⏳ Evaluation in progress — Results will be updated shortly.

Benchmark	Score
HumanEval pass@1	⏳
MMLU-Pro	⏳
GSM8K	⏳
ARC-Challenge	⏳

---

Model Details

Property	Value
Base architecture	Qwen3.5-9B
Parameters	~9B
Merge method	DARE-TIES (weight=0.5, density=0.5)
Context length	262,144 tokens
License	Apache 2.0
Training hardware	H100 80GB (Vast.ai)

---

Available Quantizations

Quantization	Size	Use case
F16	17.9 GB	Full precision / re-quantization
Q8_0	9.5 GB	Near-lossless
Q6_K	7.4 GB	High quality
Q5_K_M	6.5 GB	Recommended for quality
Q4_K_M	5.6 GB	Best balance ← start here
Q4_0	5.3 GB	Fast inference
IQ4_XS	5.2 GB	Efficient 4-bit
Q3_K_S	4.3 GB	Small footprint
IQ3_M	4.4 GB	Small + imatrix
Q2_K	3.8 GB	Minimum quality

---

Usage

With llama.cpp (recommended)

llama-server \
    -m Qwopus-GLM-9B-DualReason-Distilled-Q4_K_M.gguf \
    --ctx-size 32768 \
    --flash-attn on \
    --n-gpu-layers 99

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "rico03/Qwopus-GLM-9B-DualReason-Distilled",
    dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "rico03/Qwopus-GLM-9B-DualReason-Distilled",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Solve step by step: 2x + 5 = 13"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))

---

Source Models

All credit for the source models goes to Jackrong.

Jackrong/Qwopus3.5-9B-v3.5

A reasoning-enhanced fine-tune of Qwen3.5-9B trained with ~2x more SFT data than v3, focused on structured reasoning, tool-augmented workflows, and multi-step agentic tasks.

• HumanEval pass@1: 87.80%
• MMLU-Pro: 90.36%
• Training guide: GitHub

Jackrong/Qwen3.5-9B-GLM5.1-Distill-v1

A distilled variant of Qwen3.5-9B trained on high-quality reasoning data from a GLM-5.1 teacher model (~700x scale). Focused on structured reasoning, instruction-following, and problem decomposition.

• Training data: Jackrong/GLM-5.1-Reasoning-1M-Cleaned

---

Why DARE-TIES and Not a Frankenmerge

This model went through an extensive research process before arriving at DARE-TIES.

Step 1 — ANN Router (abandoned)

We designed a lightweight ANN router (~1M parameters) to learn optimal layer selection between the two models:

Input hidden states (both models)
         ↓
LayerRouter MLP: Linear(8320,1024) → SiLU → LayerNorm → Dropout → Linear(1024,256) → SiLU → LayerNorm → Linear(256,1)
         ↓
α ∈ [0,1] per layer → mixed = α × h_qwopus + (1-α) × h_glm

Trained on 12,297 examples from GSM8K, MMLU, ARC-Challenge, HumanEval, and IFEval with entropy regularization. Result: the router consistently collapsed to always selecting one model regardless of training objective — entropy loss or cross-entropy.

Step 2 — Cosine Similarity Analysis

To understand why, we computed cosine similarity between hidden states of both models across all 32 layers on 12,297 examples:

Layer	CosSim	Layer	CosSim	Layer	CosSim	Layer	CosSim
00	1.000	08	0.994	16	0.994	24	0.996
01	1.000	09	0.996	17	0.994	25	0.996
02	0.999	10	0.996	18	0.996	26	0.996
03	0.999	11	0.997	19	0.996	27	0.995
04	0.998	12	0.996	20	0.996	28	0.994
05	0.998	13	0.996	21	0.996	29	0.993
06	0.996	14	0.994	22	0.997	30	0.993
07	0.995	15	0.993	23	0.996	31	0.992

Conclusion: Cosine similarity of 0.992–1.000 across all layers means the two models produce nearly identical hidden states. They differ in their weights, not in the structure of their internal representations. Layer-by-layer selection adds no value — the optimal combination must happen at the weight level.

Step 3 — DARE-TIES (adopted)

DARE-TIES operates directly on model weights, interpolating the specialized knowledge of both models continuously:

tv_g   = weights_glm - weights_qwopus      # task vector
mask   = torch.rand_like(tv_g) < 0.5       # DARE dropout
tv_g   = tv_g * mask / 0.5
sign   = tv_g.sign()                        # TIES sign resolution
tv_g   = tv_g * (tv_g.sign() == sign).float()
merged = weights_qwopus + 0.5 * tv_g       # final merge

Unlike Jackrong's frankenmerge which makes a binary choice per layer, DARE-TIES produces a continuous interpolation of both models' knowledge at every weight.

---

Why No DPO/SFT

We explored DPO post-training but abandoned it after careful analysis. The primary dataset (GLM-5.1-Reasoning-1M-Cleaned) has a median chosen length of 3,225 tokens. With max_length=512, 95.8% of examples get truncated — the chosen response (with full thinking) gets cut short while the rejected (without thinking) remains intact, inverting the DPO signal. This was confirmed empirically: loss dropped from 0.51 to 0.069 in 20 steps (false convergence from corrupted signal).

Correct DPO for this class of model requires max_length ≥ 4,096 and complete long-CoT pairs — addressed in a future version.

---

References

• Rethinking Generalization in Reasoning SFT (Ren et al., 2026) — arXiv:2604.06628
• DARE — Language Model Merging by Uncertainty-Based Model Pruning
• TIES — Resolving Interference When Merging Models
• Jackrong's fine-tuning guide — GitHub

---

Acknowledgements

• Jackrong — both source models, training pipelines, datasets, and documentation
• Qwen team — Qwen3.5-9B base model
• GLM-5.1 team — teacher model used in distillation
• Kassadin88 — original GLM-5.1-1000000x dataset
• KyleHessling1 — Qwopus-GLM-18B-Merged reference benchmark

---

Citation

@misc{rico03_qwopus_glm_dualreason,
  title     = {Qwopus-GLM-9B-DualReason-Distilled},
  author    = {rico03},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/rico03/Qwopus-GLM-9B-DualReason-Distilled}
}