Sunflower-32B¶

Overview¶

Sunflower-32B is a multilingual causal language model developed by Sunbird AI, built on Qwen3-32B and fine-tuned for translation and text generation across 31 Ugandan languages plus English. It is the flagship, full-precision (BF16) model in the Sunflower series and achieves the strongest translation accuracy of any evaluated model — including Gemini 2.5 Pro and GPT-4o — on Sunbird's internal benchmark.

Task Type¶

Text Generation (LLM) — Translation, question answering, summarization, conversational.

Base Architecture¶

Qwen3-32B (base model: Qwen/Qwen3-32B → fine-tuned as Sunbird/Sunflower-32B).

Model Size¶

33B parameters, BF16 tensor type.

Quantization Format¶

None (full precision / base model). Seven quantized derivatives of this model exist on HuggingFace (llama.cpp, LM Studio, Jan, Ollama compatible) — these should be documented as separate pages, following the same pattern as Sunflower-14B-GGUF.

Languages Supported¶

31 Ugandan languages + English.

Intended Use¶

• Translation between English and Ugandan languages
• Translation between Ugandan languages
• Text generation in Ugandan languages
• Question answering in Ugandan languages

Also available via the Sunbird AI API at api.sunbird.ai.

Known Limitations¶

• Performance varies across languages depending on training data availability
• Limited evaluation beyond translation and basic Q&A
• May reflect biases present in training data
• Not suitable for critical applications (medical, legal) without human oversight
• Works best on text similar to the training distribution
• Training data includes historical texts that may contain archaic or outdated language forms; data filtering aimed to remove harmful content but cannot guarantee the complete absence of bias. Outputs should be reviewed by target-language speakers for critical applications.

Access Note — ⚠️ Gated Model¶

Sunflower-32B is a gated repository. Users must log in to HuggingFace and accept the sharing conditions before they can download or load the model. A plain from_pretrained() call will fail without prior authentication (huggingface-cli login or an access token).

How to Load / Run¶

Transformers (pipeline)¶

from transformers import pipeline

pipe = pipeline("text-generation", model="Sunbird/Sunflower-32B")
messages = [{"role": "user", "content": "Who are you?"}]
pipe(messages)

Transformers (direct, with recommended generation settings)¶

import transformers
import torch

MODEL_PATH = 'Sunbird/Sunflower-32B'
SYSTEM_MESSAGE = 'You are Sunflower, a multilingual assistant made by Sunbird AI who understands all Ugandan languages. You specialise in accurate translations, explanations, summaries and other cross-lingual tasks.'

tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH)
model = transformers.AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map='auto',
)

instruction = "Translate from Luganda to English: Wano webawaaba?"

messages = [
    {"role": "system", "content": SYSTEM_MESSAGE},
    {"role": "user", "content": instruction}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer([prompt], return_tensors="pt").to('cuda')
outputs = model.generate(
    **inputs,
    max_new_tokens=500,
    num_beams=5,
    do_sample=True,
    temperature=0.5,
)

response = tokenizer.decode(outputs[0][len(inputs['input_ids'][0]):], skip_special_tokens=True)
print(response)

vLLM (serving)¶

pip install vllm
vllm serve "Sunbird/Sunflower-32B"

# Call via OpenAI-compatible API
curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "Sunbird/Sunflower-32B",
    "messages": [{"role": "user", "content": "What is the capital of France?"}]
  }'

SGLang (serving)¶

pip install sglang
python3 -m sglang.launch_server \
    --model-path "Sunbird/Sunflower-32B" \
    --host 0.0.0.0 \
    --port 30000

Docker Model Runner¶

docker model run hf.co/Sunbird/Sunflower-32B

Note: At 33B parameters in BF16, this model requires a substantial GPU or multi-GPU setup to run at full precision. For local or low-resource deployment, consider one of the quantized variants instead.

Training Details¶

Training data (~750M characters): - Digitized books and educational materials - Radio transcripts (500+ hours transcribed) - Web data from MADLAD-400 and Common Crawl - Existing multilingual datasets (SALT, FLORES-200, MT560, TICO-19) - Dictionaries, proverbs, and cultural documents

Training procedure — 3 stages: 1. Continued pretraining — base Qwen3-32B, ~6 hours on 4× H200 GPUs, next-token prediction, DeepSpeed ZeRO-3, batch size 32,768 tokens, LR 1e-4 2. Supervised fine-tuning — ~700 instruction-response pairs, LoRA (rank 16, alpha 16), loss computed only on response tokens 3. Preference optimization — Iterative Reasoning Preference Optimization (RPO), targeting reduced hallucination/glitching, alpha 1.0

Evaluation¶

Evaluated on a custom set of 100 sentences across 20 practical scenarios (healthcare, banking, education, agriculture, etc.) spanning all 31 languages.

Comparison with other models:

⚠️ Note: the model card's Model Description states Sunflower-32B has the highest accuracy in 24 of 31 language pairs, while the Evaluation section states 25 of 31. Both figures appear verbatim on the same HF page.

Licence¶

Apache 2.0

Hugging Face¶

huggingface.co/Sunbird/Sunflower-32B

Training dataset: huggingface.co/datasets/Sunbird/salt

Last Updated¶

October 2025

Citation¶

@misc{sunflower2025,
  title={Sunflower: A New Approach To Expanding Coverage of African Languages in Large Language Models},
  author={Akera, Benjamin and Nafula, Evelyn and Yiga, Gilbert and Natukunda, Phionah and Nsumba, Solomon and Muhanguzi, Joel and Namara, Janat and Sekalala, Imran and Walukagga, Patrick and Bainomugisha, Engineer and Mwebaze, Ernest and Quinn, John},
  year={2025},
  publisher={Sunbird AI}
}

Contact: info@sunbird.ai