Mtrini-27B-Tellus
27B-class language model on Qwen3.8-27B, focused on coding, mathematics, reasoning, Arabic and Moroccan Darija.
Mtrini-27B-Tellus is a QLoRA fine-tune of Qwen3.8-27B on an 11,200-example mix of code, math, Moroccan Darija and Arabic data.
It ships as an adapter, a merged Transformers model, a full F16 GGUF and a heavily compressed IQ2_XS GGUF — with the importance matrix and calibration text used for that quantization published alongside.
ATU-1.1 score
85.73%
Specifications
- Status
- Released
- Base model
- Qwen3.8-27B (opens in new tab)
- Parameters
- ~27B
- Context length
- 4,096 tokens
- Method
- QLoRA (4-bit NF4), BF16 compute
- LoRA rank / alpha
- 32 / 64
- GGUF architecture
- qwen35 · 64 blocks
- Focus
- Coding, Mathematics, Reasoning, Arabic, Moroccan Darija
Downloads
Training data
- OpenCodeInstruct
- 3,920 (35%)
- Magicoder
- 1,120 (10%)
- OpenR1-Math
- 2,800 (25%)
- Moroccan Darija
- 2,240 (20%)
- Aya Arabic / Moroccan Arabic
- 1,120 (10%)
- Total
- 11,200
Training run
- Steps
- 700 (1 epoch)
- Final training loss
- 0.4302
- Runtime
- ~3.15 hours
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "CompiwerAI/Mtrini-27B-Tellus-Merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)Limitations
- Performance varies with prompt format, runtime, quantization and task.
- Training loss is a training metric and should not be read as a benchmark score.
- Large-model inference may need substantial GPU or CPU memory.