---
license: apache-2.0
tags:
- text-generation
- causal-language-model
- qwen
- coding
- mathematics
- reasoning
- darija
- arabic
- moroccan-arabic
- gguf
library_name: transformers
pipeline_tag: text-generation
---

# Mtrini-27B-Tellus

**Mtrini-27B-Tellus** is a CompiwerAI 27B-class language model focused on coding, mathematics, reasoning, Arabic, and Moroccan Darija.

> **CompiwerAI — Building AI For Everyone.**

## Model Overview

| Property | Value |
|---|---|
| Model | Mtrini-27B-Tellus |
| Organization | CompiwerAI |
| Base | Qwen3.8-27B |
| Parameters | ~27B |
| Context length | 4096 tokens |
| Training steps | 700 |
| Effective batch size | 16 |
| Learning rate | 0.00015 |
| LoRA rank | 32 |
| LoRA alpha | 64 |
| LoRA dropout | 0.05 |
| Training quantization | 4-bit NF4 |
| Compute dtype | BF16 |
| Optimizer | paged_adamw_8bit |
| Packing | Disabled |

## Training Data

| Dataset category | Samples | Mix |
|---|---:|---:|
| 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** | **100%** |

## Training Run

- Global step: **700**
- Final training loss: **0.4302458722250802**
- Epoch: **1**
- Runtime: approximately **3.15 hours**
- Hardware: **NVIDIA RTX PRO 6000 Blackwell Server Edition**
- Compute: **BF16**
- Method: **QLoRA / LoRA**

## Release Variants

### LoRA Adapter

The adapter-only release contains the learned LoRA weights.

`CompiwerAI/Mtrini-27B-Tellus-Adapter`

### Merged Transformers Model

The merged release contains the adapter weights merged into the compatible base model.

`CompiwerAI/Mtrini-27B-Tellus-Merged`

### F16 GGUF

Full F16 GGUF conversion for llama.cpp-compatible runtimes.

`CompiwerAI/Mtrini-27B-Tellus-GGUF`

### IQ2_XS GGUF

Highly compressed IQ2_XS GGUF release.

`CompiwerAI/Mtrini-27B-Tellus-IQ2_XS`

### IQ2_XS Imatrix

Importance matrix and calibration resources used for IQ2_XS quantization.

`CompiwerAI/Mtrini-27B-Tellus-IQ2_XS-Imatrix`

## GGUF Architecture

- Architecture: `qwen35`
- Tensor count: 851
- Transformer blocks: 64
- Embedding size: 5120
- FFN size: 17408
- Attention heads: 24
- KV heads: 4

## Intended Use

Mtrini-27B-Tellus is intended for coding assistance, mathematics, reasoning, Arabic language tasks, Moroccan Darija tasks, general text generation, and local inference experimentation.

## Limitations

Performance can vary depending on prompt format, runtime, quantization, and task.

Training loss is a training metric and should not be interpreted as a benchmark score.

## Organization

**CompiwerAI**

Building AI For Everyone.