DEV Community: Umuhoza Marie solange

# Finding the Right Model: Translating Kinyarwanda to English with Open-Source Language Models

Umuhoza Marie solange — Mon, 26 Oct 2026 14:42:37 +0000

Introduction

Language is an important part of communication, education, and access to information. However, people who speak languages with fewer digital resources may face difficulties when using artificial intelligence tools. Many language models work well with widely used languages, such as English, but their performance can be different when handling less represented languages.

Our project focuses on Kinyarwanda-to-English translation. The goal is to explore how artificial intelligence can help translate Kinyarwanda sentences into clear and meaningful English. While studying Natural Language Processing (NLP), I learned about Transformers, attention mechanisms, GPT-style models, and how large language models generate text. These concepts helped me understand how modern language models process and produce language.

In this blog, I compare three open-source language models: Qwen3-4B-Instruct-2507, Llama 3.2 3B Instruct, and SmolLM3-3B. I then select one model for further testing in my project.

1. Overview of Three Open-Source Language Models

Qwen3-4B-Instruct-2507

Qwen3-4B-Instruct-2507 was developed by the Qwen team at Alibaba Cloud. It has approximately four billion parameters and supports a context window of 262,144 tokens. Its licence is Apache 2.0, which allows users to use, modify, and distribute the model under the licence conditions.

This model is designed for instruction following, reasoning, text comprehension, mathematics, and code generation. Its developers also report improvements in multilingual capabilities. These features make it a useful candidate for testing language translation.

For our project, we can give the model a Kinyarwanda sentence and instruct it to return only the English translation. However, because it is a general-purpose language model, we must evaluate its Kinyarwanda translation quality rather than assume it will always produce accurate results.

Model card: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507

Llama 3.2 3B Instruct

Llama 3.2 3B Instruct was developed by Meta. It contains approximately 3.21 billion parameters and supports a context window of up to 128,000 tokens. It uses the Llama 3.2 Community License, which has specific conditions for use and distribution.

The model is intended for conversational assistance, summarisation, text generation, and information retrieval. Its relatively small size makes it worth considering for applications with limited computing resources.

However, Meta lists eight officially supported languages for Llama 3.2, and Kinyarwanda is not among them. This does not prove that it cannot translate Kinyarwanda, but its performance would need to be tested carefully.

Model card: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct

SmolLM3-3B

SmolLM3-3B was developed by Hugging Face. It has approximately three billion parameters and supports a context window of up to 128,000 tokens. It is released under the Apache 2.0 licence.

SmolLM3-3B is designed for instruction following, reasoning, and general text generation. Hugging Face also provides information about its training data and methods, making it useful for learning and experimentation.

Its reported language support includes English, French, Spanish, German, Italian, and Portuguese. Kinyarwanda is not included in this list. Therefore, I would need to evaluate its translation performance before considering it suitable for our project.

Model card: https://huggingface.co/HuggingFaceTB/SmolLM3-3B

2. Comparison of the Models

The following table summarises the main differences between the three models.

Feature	Qwen3-4B-Instruct-2507	Llama 3.2 3B Instruct	SmolLM3-3B
Developer	Qwen team, Alibaba Cloud	Meta	Hugging Face
Parameters	4 billion	3.21 billion	3 billion
Context window	262,144 tokens	128,000 tokens	Up to 128,000 tokens
Licence	Apache 2.0	Llama 3.2 Community License	Apache 2.0
Main strengths	Instruction following, reasoning, multilingual text tasks	Chat, summarisation, text generation	Reasoning and instruction following
Kinyarwanda support	Requires evaluation	Not officially listed	Not among reported supported languages

The comparison shows that all three models can perform general language tasks, but they differ in size, licensing conditions, and reported language support. Qwen3-4B-Instruct-2507 has the largest parameter count and context window among these three models. However, model size and context length alone do not determine translation quality.

3. The Model I Choose and Why

After comparing the models, I choose Qwen3-4B-Instruct-2507 for further testing in my Kinyarwanda-to-English translation project.

First, it is suitable for experimentation with limited computing resources. A model with four billion parameters requires less memory than many much larger language models. In 16-bit precision, its weights require approximately eight gigabytes of memory. A free Google Colab T4 GPU has 16 GB of GPU memory, so running this model is a reasonable option if memory is managed carefully. Additional memory is required for processing inputs and generating translations. If necessary, I can use quantization to reduce memory requirements.

Second, the model weights are available under the Apache 2.0 licence. I can download and run the model using Python libraries such as Hugging Face Transformers without paying for a model API or supplying an API key. I must still follow the licence and any applicable platform terms. Free Colab access also has usage limits and does not guarantee that a T4 GPU will always be available.

Third, Qwen3-4B-Instruct-2507 is designed to follow instructions. This is important because my project needs a consistent output format. For example, I can instruct the model to translate a Kinyarwanda sentence into natural English and return only the translated sentence.

For a simple example, the input could be “Ndashaka kujya ku isoko.” The expected English translation is “I want to go to the market.” This example illustrates the task format, but it does not prove that the model will translate every sentence correctly.

Finally, I will evaluate the model using a collection of Kinyarwanda sentences paired with correct English translations. I will compare its outputs with the reference translations and use evaluation measures such as BLEU or chrF. I will also manually review examples to check meaning, grammar, and whether important information has been preserved.

Qwen3 is not specifically developed only for Kinyarwanda translation, so it may make mistakes with idioms, context, or uncommon words. If its performance is not satisfactory, I can explore improving the prompts or fine-tuning the model with suitable Kinyarwanda-English parallel data.

Conclusion

Open-source language models provide opportunities to explore NLP projects without depending entirely on paid APIs. Qwen3-4B-Instruct-2507, Llama 3.2 3B Instruct, and SmolLM3-3B each offer useful capabilities, but their reported language support and licensing conditions differ.

I selected Qwen3-4B-Instruct-2507 because of its instruction-following capabilities, multilingual features, Apache 2.0 licence, and potential to run on a free Google Colab T4 GPU. My next step is to test it with real Kinyarwanda-English examples and measure the quality of its translations. This evaluation will help determine whether the model is useful for my project and what improvements may be needed.

References

Qwen Team. Qwen3-4B-Instruct-2507 Model Card. Hugging Face. https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
Meta. Llama 3.2 3B Instruct Model Card. Hugging Face. https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
Hugging Face. SmolLM3-3B Model Card. Hugging Face. https://huggingface.co/HuggingFaceTB/SmolLM3-3B
Vaswani, A., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems. https://arxiv.org/abs/1706.03762
Hugging Face. Transformers Documentation. https://huggingface.co/docs/transformers/

My Computer Vision Learning Journey

Umuhoza Marie solange — Tue, 13 Oct 2026 14:11:26 +0000

How can a computer recognize a person in a picture? How can it identify a car or detect an object in an image?
These are some of the questions I explored while learning Computer Vision.
As a beginner, I wanted to understand how computers work with images and how AI models recognize objects. I started by learning about pixels and image processing, then moved on to image classification, object detection, YOLO, and image segmentation.
Here are some of the things I learned along the way.
1. How Do Computers See Images?
The first thing I learned is that computers represent images using numbers.
A digital image is made up of small units called pixels. Each pixel contains information about the color or brightness at a particular point in the image.
One common way to represent color is RGB, which stands for Red, Green, and Blue.
For example:
Red: (255, 0, 0)
Green: (0, 255, 0)
Blue: (0, 0, 255)
Black: (0, 0, 0)
White: (255, 255, 255)
A color image usually has three channels, one for each color. This means a 200 × 300 image contains 180,000 color values.
Understanding pixels helped me see that a computer does not see a photograph exactly as a human does. It processes the numerical information that represents the image.

2. Learning Basic Image Processing
After learning about pixels, I explored different ways of changing images using Python.
Grayscale: Converts a color image into shades of gray. It can make some image-processing tasks simpler, but it removes color information.
Blurring: Makes an image less sharp. It can reduce some noise, but too much blur can hide important details.
Sharpening: Makes edges and details appear clearer, although excessive sharpening can affect image quality.
Edge detection: Helps identify boundaries where the brightness changes. These boundaries can reveal the shapes of objects.
I also learned about the Sobel operator, which uses small filters to detect changes in image brightness.
This introduced me to the idea of convolution. I learned that filters can be applied to image pixels to identify patterns, while Convolutional Neural Networks can learn useful filters from data.
These concepts helped me understand some of the basic ideas behind image analysis.

** 3. What Is Image Classification?**
Image classification is the process of predicting what an image contains. For example, a model might classify an image as a dog, a car, or a computer screen.
While learning about image classification, I used a pretrained model to analyze an image. The model returned several possible labels, each with a different score:

Website: 87.5%
Menu: 8.4%
Screen: 2.2%
Television: 0.7%
Brass: 0.4% The model predicted “website” with the highest score of 87.5%. This helped me understand that a classification model can provide several possible predictions, ranked by their scores. However, the highest score does not always mean the prediction is correct. I also learned that image classification generally predicts what an image contains without showing the exact location of an object. This is one difference between image classification and object detection. Object detection can identify objects and show their locations using bounding boxes.

4. What Is Object Detection?
Object detection identifies objects in an image and shows where they are located.
It usually provides three things:
Label: The name of the detected object.
Bounding box: A rectangle showing where the object is located.
Confidence score: A score associated with the prediction.
For example, a model could identify two people in a photograph and draw a box around each person.
This is different from classification because object detection can identify multiple objects and show their locations in the same image.

5. What Is YOLO?
YOLO stands for You Only Look Once. It is a family of AI models used for object detection.
I learned that YOLO can identify objects, predict their locations, and provide confidence scores.
The general process involves giving an image to the model, processing its visual features, predicting possible objects, and filtering the results.
The final output can display bounding boxes and labels on the image.
I also learned that object detection models can produce duplicate or incorrect predictions. This is why the results need to be checked rather than accepted automatically.

6. Understanding Confidence Thresholds
A confidence threshold controls which predictions an object detection model keeps.
For example, setting a threshold to 0.10 means using a score of 10% as the filtering threshold in a typical detection setup.
A lower threshold can allow more predictions, including incorrect ones. A higher threshold filters out more low-scoring predictions, but it can also remove real objects.
While exploring object detection, I observed predictions such as cake, toothbrush, and scissors, even though those labels did not appear to match the screenshot I was examining.
This helped me understand that models can make mistakes.
I learned that a confidence score is not a guarantee of correctness. We need to examine the image and evaluate the predictions to understand how well a model is performing.

7. What Is Image Segmentation?
Image segmentation is another Computer Vision task. It identifies regions or pixels belonging to particular categories or objects.
I learned about three types:
Semantic segmentation: Assigns a category to each pixel, such as road, building, or person.
Instance segmentation: Separates individual objects, even when they belong to the same category.
Panoptic segmentation: Combines semantic and instance segmentation.
The main difference is that object detection uses bounding boxes, while segmentation can provide a more detailed representation of an object's shape.
Learning about segmentation helped me understand that Computer Vision includes more than image classification and object detection.

8. How Do Image Changes Affect Predictions?
I also explored how changing an image can affect what a model predicts.
Some of the changes I studied included:
Making an image darker.
Rotating an image upside down.
Reducing its resolution.
Applying blur.
I learned that these transformations can change the visual information available to a model.
For example, a blurred image may hide important details, while a low-resolution image may lose information about small objects. Making an image larger again does not restore details that were lost when reducing its resolution.
I also learned that experiments should be carried out carefully. To understand the effect of a transformation, it is important to compare the results using suitable images and consistent settings.

9. Why Do Computer Vision Models Make Mistakes?
One important thing I learned is that a model can run successfully and still produce an incorrect result.
Some reasons include:
Poor lighting or image quality.
Objects that are too small or partly hidden.
Similar-looking objects.
Limitations in the model.
Differences between the training images and the images being analyzed.
This helped me understand why we should not depend only on a model's output. We need to check its predictions and test its performance.

10. What Is Transfer Learning?
Transfer learning involves using a model that has already been trained and adapting it to another task.
Instead of training a model from the beginning, we can start with a pretrained model and fine-tune it using suitable data.
For example, a pretrained model could be adapted to recognize certain crop diseases using labelled images of local crops.

I learned that the quality and relevance of the training data matter. A model should also be tested on separate images to see how well it works on examples it has not seen during training.

This is especially important when developing Computer Vision applications for different environments, including Rwanda.

11. Applications of Computer Vision
Computer Vision can be useful in many areas.
Agriculture: Identifying plant diseases and monitoring crop growth.
Healthcare: Supporting the analysis of medical images under appropriate professional supervision.
Transportation: Detecting vehicles, pedestrians, and road signs.
Environment: Monitoring vegetation, wildlife, and changes in land use.
Accessibility: Helping people with visual impairments understand objects and text around them.

These are examples of how Computer Vision could help solve real-world problems. However, each application needs to be tested carefully to make sure it works reliably in its intended environment.

12. What I Have Learned
Studying Computer Vision has helped me understand several important concepts:
Images are represented using pixels and numerical values.
Image processing techniques change how images look.
Image classification predicts what an image contains.
Object detection identifies objects and their locations.
YOLO is used for efficient object detection.
Confidence thresholds affect which predictions are retained.
Image segmentation provides information about individual pixels and object shapes.
Changes in image quality can affect model predictions.
Models can make mistakes even when the code runs correctly.
Transfer learning can help adapt pretrained models to new tasks.

What I Want to Learn Next

I want to continue learning about Convolutional Neural Networks, model training, and how to evaluate Computer Vision models using metrics such as precision, recall, and mean Average Precision.

I also want to improve my Python skills and learn how to prepare datasets and fine-tune models for practical applications.

Conclusion

My learning journey has helped me understand the basics of Computer Vision and how AI models work with images.
From pixels and image processing to classification, YOLO, and segmentation, each topic has helped me understand a different part of the field.
I am still a beginner, but I now have a better understanding of how models make predictions and why those predictions are not always correct.
I look forward to learning more and using these skills to explore practical problems in Rwanda and beyond.