Meet Sunflower, Uganda’s pioneering multilingual AI model, launched just days ago on October 10, 2025, at the AfrilangAI2025 conference in Kampala. Developed by the non-profit Sunbird AI, this open-source powerhouse isn’t just another large language model (LLM)—it’s a cultural bridge, a tool for empowerment, and a symbol of resilience.
Named after the sunflower for its themes of vision, growth, and inclusion, Sunflower turns the spotlight on Uganda’s 40+ indigenous languages, making AI accessible, relevant, and truly African.
As Dr. Aminah Zawedde, Uganda’s Minister of State for Science, Technology, and Innovation, proclaimed at the launch: “This is more than technology; it’s a cultural milestone that removes language barriers for our young innovators.”
In an era dominated by English-centric AIs like ChatGPT, Sunflower flips the script, prioritizing low-resource languages from the Bantu, Nilotic, and Central Sudanic families. Let’s dive deep into what makes this model a game-changer for African tech.
Sunbird AI’s mission
Sunbird AI, headquartered in Kampala, Uganda, is a non-profit research lab laser-focused on practical AI for African contexts. Founded to tackle real-world challenges in under-resourced environments, the organization collaborates with NGOs, universities, SMEs, and community groups across speech and language tech, geospatial sensing, and public-sector tools. Their ethos? Build reliable, locally tuned AI that delivers societal benefits while being transparent about biases and limitations.
Sunflower’s development embodies this: it’s not scraped from the web but drawn from authentic, community-sourced data like transcribed radio broadcasts, digitized out-of-print books, school materials, and cultural archives. Partners such as Makerere University and the Cross-Cultural Foundation of Uganda ensured the data reflects genuine voices—verified by local teachers and translators. This “build with people” approach contrasts sharply with Big Tech’s top-down models, emphasizing open-source releases under Apache 2.0 licenses for reproducibility and ethical use.
Sunflower’s core
At its heart, Sunflower comprises two fine-tuned large language models (LLMs) built on the robust Qwen 3 architecture: Sunflower-14B and Sunflower-32B. These aren’t off-the-shelf imports slapped with a regional label; they’re the result of meticulous adaptation tailored to Uganda’s rich linguistic tapestry, supporting seamless communication in 31 indigenous languages—spanning the Bantu family (like Luganda and Runyankore), the Nilotic group (such as Acholi and Ateso), and the Central Sudanic branch (including Lango and Madi)—alongside English. This broad coverage enables Sunflower to handle a wide array of tasks with cultural sensitivity, from translating everyday greetings to generating contextually appropriate stories or answering questions about local folklore.
The Sunflower-14B variant packs approximately 14.8 billion parameters, making it an efficient causal language model ideal for core functions like real-time translation, text generation, and question-answering. It’s particularly suited for resource-conscious deployments, where speed and accessibility matter most. Developers often opt for its W8A8 quantization—using 8-bit weights and activations—which dramatically reduces memory footprint without sacrificing much accuracy. For instance, this can halve the model’s FP16 requirements from around 28GB to about 14GB, allowing it to run smoothly on NVIDIA GPUs like the V100 or A100 with just 12-16GB of VRAM and at least 32GB of system RAM. This setup ensures low-latency responses, even in environments with limited compute power, such as university labs or small-scale NGOs in rural Uganda.
Stepping up in scale and sophistication is the Sunflower-32B model, with roughly 32 billion parameters, which excels in more demanding scenarios requiring deeper reasoning or nuanced cultural handling. It builds on the same multilingual foundation but delivers enhanced performance on complex, multi-step tasks—like summarizing lengthy oral histories or generating interactive dialogues that weave in regional idioms. Like its smaller sibling, it supports W8A8 quantization for broad compatibility, but it also offers FP4A16 options for even greater efficiency on supported hardware. Hardware needs mirror the 14B’s but lean toward the higher end, with 16GB or more VRAM recommended to unlock its full potential without throttling. Both models leverage a structured chat template, incorporating system and user roles to guide interactions, and are optimized for INT8 tensor cores on NVIDIA GPUs, further minimizing delays in real-world applications.
To get hands-on with these models, developers can download them directly from Hugging Face repositories, where quantized versions are readily available for quick integration. Using a library like vLLM, the process is straightforward: first, load the chosen model variant (say, the quantized Sunflower-14B-W8A8) with an eager execution flag to bypass certain optimizations for compatibility. Then, define sampling parameters—such as a temperature of 0.7 for balanced creativity, a top-p value of 0.9 to focus on high-probability tokens, and a max token limit of 1,000 to control output length. Finally, feed in your prompt, like requesting a translation from English to Luganda, and generate the response. This pipeline not only simplifies deployment but also allows for easy experimentation, such as fine-tuning on custom datasets to extend Sunflower’s capabilities to niche dialects or domains.
Quantized versions across both models slash memory demands by up to 50%, making inference faster and more accessible on modest hardware— a critical feature for scaling AI in power-variable regions like East Africa. This chat-templated setup (with system/user roles) optimizes for INT8 tensor cores on NVIDIA GPUs, ensuring low-latency responses even in resource-constrained settings.
Features that speak your language
Sunflower’s magic lies in its versatility. Beyond basic chat, it excels at:
- Translation and Summarization: Seamlessly convert news articles from English to Ateso or condense stories in Runyankore.
- Question-Answering: Query cultural proverbs or agricultural tips in your mother tongue.
- Creative Generation: Craft poems or dialogues infused with local idioms.
The user-friendly web interface at sunflower.sunbird.ai supports 20+ languages via a WhatsApp-like messenger. Just send a query, agree to privacy terms, and pick a category with pre-loaded prompts. For voice-first users—like farmers dictating crop queries in Acholi—integrations with WhatsApp enable transcription and translation on the fly.
It’s designed for the masses: no PhD required. Early testers rave about its natural tone, capturing the rhythm of Ugandan speech without the stiffness of global models.
From radio waves to AI training
Crafting Sunflower wasn’t a sprint—it was a months-long quest to diagnose why off-the-shelf LLMs flop on African languages. Sunbird’s team pinpointed issues like sparse training data and cultural mismatches, then built a reproducible pipeline: data curation from non-digital sources (e.g., 500+ hours of Simba FM radio), ethical sourcing, and targeted fine-tuning.
Key datasets include endangered folklore from Backup Uganda and educational texts from local publishers. This bottom-up method fosters “knowledge transfer” between related languages, boosting performance without massive compute. The full blueprint is all in the arXiv preprint: Sunflower: A Multilingual LLM for Under-Resourced African Languages. Developers can adapt it for other regions, scaling Sunflower’s sunflower seeds across Africa’s linguistic garden.
