MiaBee-VL-4B-TR (beta)

Turkish-adapted Qwen3-VL-4B-Instruct with vision retained. GGUF Q4_K_M.

Beta. Experimental Turkish adaptation, not a replacement for the base model.

Training

QLoRA r=32, 1 epoch, 405 steps, vision tower frozen. Data: 3,485 Turkish instruction samples + 3,000 from merve/turkish_instructions.

Evaluation (self-run, not official)

Test This model Base
TurkishMMLU, 5-shot, 900 q 55.8% 54.3%
Turkish instruction-following, 20 tasks 65% 70%
ScreenSpot-v2 Desktop, 334 screenshots 62.9% โ€”
Tool calling, 10 tasks 9/10 โ€”

Differences on the first two rows are within noise. This model does not outperform the base.

Usage

llama-server -m MiaBee-VL-4B-TR-Q4_K_M.gguf --mmproj mmproj-Q8_0.gguf \
  -c 4096 -np 1 --jinja -ctk q8_0 -ctv q8_0

~400 MB RAM text, ~900 MB with vision, ~45 tok/s.

Limitations

Hallucinates numbers; can loop on open-ended prompts; does not retry after tool errors; mmproj must be Q8_0 (Q4_0 is broken).

License

Apache 2.0. Derivative of Qwen3-VL-4B-Instruct.

Downloads last month
51
GGUF
Model size
4B params
Architecture
qwen3vl
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for mailabs/MiaBee-VL-4B-TR

Quantized
(120)
this model