QLoRA fine-tune of Qwen3.5-35B-A3B (MoE, 3B active parameters) with Claude Opus 4.6 reasoning distillation. Training recipe adapted from Jackrong's Qwopus3.5-27B-v3.

This is the full BF16 safetensors model. For GGUF quantizations (Q4, Q5, Q6, Q8), see samuelcardillo/Qwopus-MoE-35B-A3B-GGUF.

Credits

This model is based on the work of Jackrong and his Qwopus3.5-27B-v3 training methodology — same datasets, same philosophy, adapted for the MoE architecture. See his complete training guide.

Model Details

Property	Value
Base Model	Qwen/Qwen3.5-35B-A3B
Architecture	Mixture of Experts (MoE)
Total Parameters	~35B
Active Parameters	~3B per token
Precision	BF16

Training Details

Parameter	Value
Method	QLoRA (4-bit base + LoRA in BF16)
Framework	Unsloth 2026.4.2 + TRL
LoRA Rank	32
LoRA Alpha	32
LoRA Targets	q_proj, k_proj, v_proj, o_proj
Trainable Parameters	6,881,280 (0.02%)
Epochs	2
Final Loss	0.5517
GPU	NVIDIA RTX PRO 6000 Blackwell (96GB)
Training Time	~3.5 hours

Datasets (3,209 examples)

Dataset	Examples
nohurry/Opus-4.6-Reasoning-3000x-filtered	2,326
Jackrong/Qwen3.5-reasoning-700x	633
Roman1111111/claude-opus-4.6-10000x	~250

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "samuelcardillo/Qwopus-MoE-35B-A3B",
    torch_dtype="bfloat16",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "samuelcardillo/Qwopus-MoE-35B-A3B",
    trust_remote_code=True,
)

Acknowledgements

Jackrong — Training methodology and the Qwopus concept
Unsloth — QLoRA training framework
Qwen — Base model