Qwen3.5-text-9B-bnb-4bit

BNB NF4 4-bit quantization of techwithsergiu/Qwen3.5-text-9B — a text-only derivative of Qwen/Qwen3.5-9B.

No visual tower — text input only. This is the recommended base for Unsloth LoRA text fine-tuning: smaller VRAM footprint, no visual-dependency complexity, cleaner adapter targeting.

Inference has been verified. LoRA fine-tuning docs are pending — see Fine-tuning section below.

What was changed from the original Qwen3.5-9B

• Visual tower removed (same as Qwen3.5-text-9B)
• Text backbone quantized to BNB NF4 double-quant (bnb_4bit_quant_type=nf4, bnb_4bit_compute_dtype=bfloat16)
• lm_head.weight kept at bf16 for output quality / stability

Model family

Model	Type	Base model
Qwen/Qwen3.5-9B	f16 · VLM · source	—
techwithsergiu/Qwen3.5-9B-bnb-4bit	BNB NF4 · VLM	Qwen/Qwen3.5-9B
techwithsergiu/Qwen3.5-text-9B	bf16 · text-only	Qwen/Qwen3.5-9B
techwithsergiu/Qwen3.5-text-9B-bnb-4bit	BNB NF4 · text-only	Qwen3.5-text-9B
techwithsergiu/Qwen3.5-text-9B-GGUF	GGUF quants	Qwen3.5-text-9B

The visual tower scales with model size (~0.19 GB for 0.8B, ~0.62 GB for 2B/4B, ~0.85 GB for 9B). BNB text-only models are roughly 34% of the original f16 size (4B example: 9.32 GB → 3.12 GB).

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "techwithsergiu/Qwen3.5-text-9B-bnb-4bit"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "What is the capital of Romania?"}]

# Thinking OFF — direct answer
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(response)

# Thinking ON — chain-of-thought before the answer
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(response)

Fine-tuning

This is the primary training target for text LoRA fine-tuning.

Training pipeline: github.com/techwithsergiu/qwen-qlora-train

QLoRA (Unsloth + TRL + PEFT) · rank 16–64 · validated on RTX 3070 8 GB

# Quick start — install
pip install "unsloth[cu124-torch260] @ git+https://github.com/unslothai/unsloth.git"
pip install git+https://github.com/techwithsergiu/qwen-qlora-train.git

# Train with a ready-made config
qlora-train configs/qwen35/0.8b.yaml   # or 2b / 4b

After training, test the adapter without merging:

qlora-infer \
  --model   techwithsergiu/Qwen3.5-text-9B-bnb-4bit \
  --adapter adapters/<run_name>

For VLM (image + text) fine-tuning of the full model, see: unsloth.ai/docs/models/qwen3.5/fine-tune

Pipeline diagram

Conversion

Converted using qwen35-toolkit — a Python toolkit for BNB quantization, visual tower removal, verification and HF Hub publishing of Qwen3.5 models.

---

Acknowledgements

Based on Qwen/Qwen3.5-9B by the Qwen Team. If you use this model in research, please cite the original:

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}