Mtrini-SVL-2.0
A Qwen3.5-9B fine-tune that thinks before answering, understands images and speaks Moroccan Darija, with a 262,144-token context.
Mtrini-SVL-2.0 is an AI assistant built by Compiwer AI, fine-tuned from Qwen3.5-9B. It thinks before answering, understands images, supports a 262,144-token context (extensible to about 1M) and speaks Moroccan Darija.
It was trained in rounds of LoRA fine-tuning, each merged before the next, followed by a round of Gated Self-Training: on hard math, coding problems checked by their tests and hard science, the model's own right answers were preferred over its own wrong ones of similar length.
The card marks the release as experimental. Compiwer AI has since run it on four public benchmarks in full (MMLU-Pro on a 1,008-question sample), alongside the earlier quick checks.
Benchmarks
| Benchmark | Score | ||
|---|---|---|---|
| MATH-500 | Accuracy | 500 problems | 92.8% |
| HumanEval+ | pass@1 | 164 problems | 89.6% |
| IFEval | Prompt-level strict | 541 prompts | 84.7% |
| MMLU-Pro | Accuracy | 1,008 questions, 72 per subject | 73.4% |
| AIME 2025 | Accuracy, average of 8 tries | 30 problems | 61.7% |
| LiveCodeBench v6 | pass@1 | 2025 problems | 48.8% |
| GPQA Diamond | pass@1 | Diamond set | 62.1% |
| DarijaMMLU | Accuracy | 400 questions | 64.2% |
How it compares
| Model | AIME 2025 | LiveCodeBench v6 | GPQA Diamond | DarijaMMLU | Identity |
|---|---|---|---|---|---|
| Mtrini-SVL-2.0 | 61.7 | 48.8 | 62.1 | 64.2 | 93.8 |
| Qwen3.5-9B | – | – | – | 34.5 | 0.0 |
| Before Gated Self-Training | 62.1 | 47.5 | 60.6 | 64.0 | 93.8 |
Every row was run by Compiwer AI with the same questions and settings, as listed on the model card. They are quick checks on small samples, not official benchmark runs.
* Identity is the share of answers in which the model says it is Mtrini, built by Compiwer AI.
The best published score in each column is in bold.
MMLU-Pro by category
| Category | Questions | Accuracy |
|---|---|---|
| Economics | 72 | 90.3% |
| Biology | 72 | 86.1% |
| Math | 72 | 86.1% |
| Physics | 72 | 83.3% |
| Psychology | 72 | 80.6% |
| Business | 72 | 79.2% |
| Chemistry | 72 | 75.0% |
| Computer science | 72 | 73.6% |
| Other | 72 | 73.6% |
| Philosophy | 72 | 73.6% |
| Health | 72 | 68.1% |
| History | 72 | 61.1% |
| Engineering | 72 | 48.6% |
| Law | 72 | 48.6% |
Specifications
- Status
- Experimental
- Base model
- Qwen3.5-9B (opens in new tab)
- Context
- 262,144 tokens, extensible to about 1M
- Thinking
- On by default; pass enable_thinking=False to the chat template for direct answers
- Formats
- Merged Transformers weights; GGUF quants by mradermacher
- License
- Apache-2.0
- Languages
- Moroccan Darija, Arabic, English
- Focus
- Reasoning, Moroccan Darija, Vision-language
Downloads
Training data
- Round 1
- LoRA SFT: 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
- Round 2
- A short pass on new Codeforces problems and more identity answers
- Round 3
- 182 steps × 16 examples (4,023 rows): OpenThoughts3 math, code and science reasoning, more Darija, identity answers and replay
- Round 4
- 123 steps × 16 examples: a new slice of OpenThoughts3 reasoning, more Darija and identity answers
- Round 7
- Gated Self-Training with preference pairs (5 DPO steps): its own right answers preferred over its own wrong ones, plus identity pairs
Usage
ollama run hf.co/mradermacher/Mtrini-SVL-2.0-GGUF:Q4_K_MLimitations
- Experimental. Check the base model’s card for its limitations, and verify important answers.
- Scores are Compiwer AI’s own runs and have not been independently re-scored. AIME, LiveCodeBench, GPQA and DarijaMMLU are quick checks on small samples.
- Law and engineering are its weakest MMLU-Pro subjects (48.6%). Check answers in those areas carefully.