LLM Research Blog: The Model that Fits our Project
Our project and its limits In our NLP course, my group (Group 1) took apart GPT-2 small , OpenAI's 124-million-parameter model from 2019, to understand how it predicts the next word . Each of us studied one part: tokens, position embedding, attention, the MLP, and mine, layer norm . We tested every
In our NLP course, my group (Group 1) took apart GPT-2 small, OpenAI's 124-million-parameter model from 2019, to understand how it predicts the next word. Each of us studied one part: tokens, position embedding, attention, the MLP, and mine, layer norm. We tested everything on the same sentence in English and Kinyarwanda, and built a small Gradio chat demo.
So I researched six open-source LLMs to choose the one our team would use next.
Qwen3-4B-Instruct-2507 (Alibaba). A 4.0B-parameter chat model with a 262,144-token context, under Apache 2.0. It is the model from our lecture lab and scores 83.4 on IFEval, a test of how well a model follows instructions.
Gemma 3 4B (Google). A 4B model with a 128K context and over 140 languages. It is multimodal: it can read an image plus text. It uses the Gemma licence, and you must log in and accept Google's terms to download it.
Llama 3.2 3B Instruct (Meta). A 3.21B model with a 128K context, built for small devices. It scores 77.4 on IFEval and officially supports eight languages, none African. Its download needs your contact details and Meta's approval.
Phi-4-mini-instruct (Microsoft). A 3.8B model with a 128K context, under the MIT licence with no gate. It is strongest at math and logic (88.6 on GSM8K) and code.
SmolLM2-1.7B-Instruct (Hugging Face). The smallest: 1.7B parameters, an 8K context, Apache 2.0. It is fast, but scores only 56.7 on IFEval and works mainly in English.
Sunflower-14B (Sunbird AI, Uganda). A 15B model built on Qwen3-14B and trained on 31 African languages, including Kinyarwanda, under Apache 2.0. It is the only one that lists Kinyarwanda, but it is too big for a free Colab GPU unless compressed.
Memory = parameters × 2 bytes. "–" = no score on the model card.
Does it run on a free Colab T4? Yes. It needs about 8 GB in 16-bit, half of the T4's 16 GB. The T4 has no bfloat16, so we load it in float16. Sunflower would need about 30 GB.
Is it free, with no API key or account? Yes. It is Apache 2.0 and downloads without logging in. Gemma and Llama make every teammate create an account and accept a licence.
Does it follow instructions? It has the best IFEval score we found (83.4), and it answers directly without a long "thinking" section. That is exactly what our chat demo lacked.
How fast is it, and how much memory? It is about 32× bigger than GPT-2, so it will be slower and use 8 GB instead of 0.25 GB. We haven't measured speed yet; our first test is to time tokens per second on the T4. If it is too slow, Phi-4-mini is our backup.
What we give up: Gemma 3 can read images and covers more languages, but its licence gate slows down a six-person team, and our project is about text. Sunflower knows Kinyarwanda, but we can't run it at full size on our free GPU.
A model card that says "multilingual" doesn't mean it knows our language: most of these models handle several languages, but only one lists Kinyarwanda. Size alone isn't the answer either. A 4B model that fits our GPU, downloads freely and follows instructions is more useful to us than a bigger model we can't run. Our next step is to load Qwen3 in Colab, time it, and repeat our GPT-2 experiments on it.