Qwen3.5-text-9B-GGUF

GGUF quants of techwithsergiu/Qwen3.5-text-9B — the text-only bf16 derivative of Qwen/Qwen3.5-9B.

The visual tower has been removed before conversion. All text-backbone weights are identical to the original — no retraining, no weight changes, no quality loss for text tasks.

Quants

File	Type	Size	Notes
Qwen3.5-text-9B-Q8_0.gguf	Q8_0	~53% of f16	near-lossless — for high-quality inference
Qwen3.5-text-9B-Q6_K.gguf	Q6_K	~41% of f16	excellent quality, good balance with f16
Qwen3.5-text-9B-Q5_K_M.gguf	Q5_K_M	~37% of f16	very good quality, smaller than Q6
Qwen3.5-text-9B-Q4_K_M.gguf	Q4_K_M	~31% of f16	✅ recommended — best size/quality balance
Qwen3.5-text-9B-Q4_K_S.gguf	Q4_K_S	~30% of f16	optional — slightly smaller, slightly lower quality

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 GGUF repo is derived from the text-only f16 model — same weights, different container format. base_model points to the f16 text variant to keep the VLM and text lineages distinct on the Hub.

Inference

llama.cpp

./llama.cpp/build/bin/llama-cli \
    -m Qwen3.5-text-9B-Q4_K_M.gguf \
    -p "What is the capital of Romania?" \
    -n 256

LM Studio

Load any .gguf file from this repo directly in LM Studio. Recommended quant: Q4_K_M.

Thinking mode

Qwen3.5 supports an optional chain-of-thought <think> block before the answer. Thinking is enabled by default in llama.cpp.

Note: --chat-template-kwargs '{"enable_thinking":...}' is deprecated — do not use. Known issue: --reasoning off is accepted but does not actually disable thinking. Workaround: use --reasoning-budget 0 — this reliably disables the <think> block. Track the bug at llama.cpp issues.

# Thinking OFF — direct answer (workaround: --reasoning-budget 0)
./llama.cpp/build/bin/llama-cli \
    -m Qwen3.5-text-9B-Q4_K_M.gguf \
    --reasoning-budget 0 \
    -p "What is the capital of Romania?" \
    -n 256

# Thinking ON — default, no flag needed
./llama.cpp/build/bin/llama-cli \
    -m Qwen3.5-text-9B-Q4_K_M.gguf \
    -p "What is 17 × 34?" \
    -n 1024

Pipeline diagram

From fine-tuned adapter to GGUF

If you have a LoRA adapter trained with qwen-qlora-train, merge it first, then convert to GGUF:

# 1. Merge adapter into f16 weights
qlora-merge \
  --base  Qwen/Qwen3.5-9B \
  --adapter adapters/<run_name> \
  --output merged/qwen35-text-9B-sft-f16

# 2. Convert merged model to GGUF  (requires llama.cpp)
python llama.cpp/convert_hf_to_gguf.py merged/qwen35-text-9B-sft-f16 \
    --outtype f16 \
    --outfile merged/qwen35-text-9B-sft-F16.gguf

# 3. Quantize
./llama.cpp/build/bin/llama-quantize \
    merged/qwen35-text-9B-sft-F16.gguf \
    merged/qwen35-text-9B-sft-Q4_K_M.gguf \
    Q4_K_M

Full post-training workflow is documented in qwen-qlora-train → Post-merge workflow.

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}
}