Mtrini-SVL-1.0
The first SVL release: a LoRA adapter on Qwen3-VL-8B-Instruct for coding and math that keeps the base model's vision.
Mtrini (متربّي) is Darija for “trained” or “well-fit”. SVL-1.0 is a small LoRA adapter that nudges Qwen3-VL-8B-Instruct toward coding and mathematical reasoning while keeping its ability to read images.
A merged, Q8_0-quantized GGUF packages the adapter, base model and vision components into a single ~8.7 GB file for llama.cpp — no Python and no separate projector file required.
Specifications
- Status
- Released
- Base model
- Qwen3-VL-8B-Instruct (opens in new tab)
- Base parameters
- 8.81B
- Method
- LoRA, rank 16
- Target modules
- Attention + MLP projections
- Training context
- 2,048 tokens
- Max steps
- 1,000
- Focus
- Coding, Mathematics, Vision
Downloads
Training data
- Bespoke-Stratos-17k
- Coding & problem-solving
- OpenR1-Math-220k
- Mathematical reasoning
Usage
llama-mtmd-cli \
-m Mtrini-SVL-1.0-Q8_0.gguf \
-p "Describe this image." \
--image path/to/your/image.jpgLimitations
- Trained on a relatively short schedule (1,000 steps) with a focused data mix.
- Inherits the strengths and limitations of the base model.
- Test on your own use case before relying on it.