---
base_model: Qwen/Qwen3.5-9B
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
language:
- ary
- ar
- en
tags: [compiwer-ai, mtrini, svl, reasoning, moroccan-darija, qwen3_5]
datasets: [simplescaling/s1K-1.1, GAIR/LIMO, open-r1/codeforces-cots, open-thoughts/OpenThoughts3-1.2M, MBZUAI-Paris/Darija-SFT-Mixture]
---

# Mtrini-SVL-2.0

Mtrini-SVL-2.0 is an AI assistant built by **Compiwer AI**, fine-tuned from [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B).
It thinks before answering, understands images, supports a 262,144-token context (extensible to about 1M), and speaks Moroccan Darija.

## Results

Quick checks run on the base model and on this model with the same questions and settings. These are small samples, not official benchmark runs.

| | Qwen3.5-9B | Before | After |
|---|---|---|---|
| AIME 2025 (30 problems, avg of 8 tries) | – | 62.1% | 61.7% |
| LiveCodeBench v6 (2025 problems, pass@1) | – | 47.5% | 48.8% |
| GPQA Diamond (pass@1) | – | 60.6% | 62.1% |
| DarijaMMLU (400 questions) | 34.5% | 64.0% | 64.2% |
| Identity (says it's Mtrini by Compiwer AI) | 0.0% | 93.8% | 93.8% |

## Training

Three rounds of LoRA SFT, each merged before the next:

1. Math and science reasoning (s1K-1.1, LIMO), Python reasoning (open-r1/codeforces-cots), Moroccan Darija (MBZUAI-Paris/Darija-SFT-Mixture), identity answers, and replay of the base model's own verified-correct reasoning.
2. A short pass on new Codeforces problems and more identity answers.
3. A larger pass (182 steps × 16 examples, 4,023 rows): OpenThoughts3 math, code and science reasoning, more Darija, identity answers and replay.

4. Round 4 (123 steps × 16 examples): a new slice of OpenThoughts3 math, code and science reasoning, more Darija and identity answers.

5. Round 7, Gated Self-Training with preference pairs (5 DPO steps): on hard math, coding problems checked by their tests, and hard science, its own right answers preferred over its own wrong ones of similar length, plus identity pairs.

## Usage

Use it like Qwen3.5-9B. Thinking is on by default; pass `enable_thinking=False` to the chat template for direct answers.

## Run it on your computer

GGUF versions for llama.cpp, LM Studio and Ollama, made by [mradermacher](https://huggingface.co/mradermacher) (thank you!), including the vision projector (`mmproj`) for images:

- [mradermacher/Mtrini-SVL-2.0-GGUF](https://huggingface.co/mradermacher/Mtrini-SVL-2.0-GGUF)
- [mradermacher/Mtrini-SVL-2.0-i1-GGUF](https://huggingface.co/mradermacher/Mtrini-SVL-2.0-i1-GGUF) (imatrix quants)

```bash
ollama run hf.co/mradermacher/Mtrini-SVL-2.0-GGUF:Q4_K_M
```

In LM Studio, search for **Mtrini-SVL-2.0**. Q4_K_M (5.6 GB) is a good default.

## Status

Experimental. Check the base model's card for limitations, and verify important answers.