Sunflower-Gemma4-E2B

[!WARNING] This is a preview release. The model weights will change between now and the stable release. Benchmark numbers, prompt behaviour, and output quality may all shift. Please don't pin production systems to this checkpoint — treat it as a preview for evaluation and feedback.

Sunflower-Gemma4-E2B is a multimodal adaptation of Google's Gemma 4 E2B IT optimised to understand text and speech across 69 African languages. The model is tuned for text translation, speech transcription, instruction-following, and multi-turn chat. It supports text, image, and audio inputs (up to 30s of 16kHz audio), generating text output.

Try it

▶ Open the Colab Demo Notebook

Supported Languages

Coverage varies considerably by language, roughly tracking the amount of training data available.

Translation & Speech Evaluation Covers: Acholi, Afrikaans, Akan, Alur, Amharic, Bambara, Bari, Bemba, Berber, Chichewa, Dagaare, Dagbani, Dinka, Ewe, French, Fulani, Hausa, Igbo, Ik, Ikposo, Jopadhola, Kabyle, Kakwa, Kalenjin, Kanuri, Karamojong, Kikuyu, Kinyarwanda, Kirundi, Kumam, Kupsabiny, Kwamba, Lango, Lendu, Lingala, Lubwisi, Luganda, Lugbara, Lugungu, Lugwere, Luhya, Lumasaba, Lunyole, Luo, Lusoga, Ma'di, Malagasy, Ndebele, Nigerian Pidgin, Oromo, Pokot, Rukiga, Rukonjo, Runyankole, Runyoro, Ruruuli, Rutooro, Samia, Shona, Somali, Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba, and Zulu.

Note: A language's presence in this list means it was evaluated; it does not imply equal quality across all languages. See evaluation metrics below for details on performance.

Inference

Sunflower-Gemma4-E2B follows standard Transformers patterns, but expects specific prompt formats for translation and speech transcription. Like Qwen, the base model supports a reasoning phase, but we recommend setting enable_thinking=False for standard translation tasks.

Text Translation

Translation expects a fixed instruction template: Translate from English to {language name}: {text}

messages = [
    {"role": "system", "content": "You are Sunflower, a helpful assistant made by Sunbird AI who knows many African languages."},
    {"role": "user", "content": "Translate from English to Luganda: Where is the nearest hospital?"}
]

Tip: Use do_sample=False (greedy decoding) and temperature=0 for translation tasks.

Audio Transcription

The model was trained with prompts that explicitly name the spoken language. Convert audio to a 16 kHz mono waveform and keep it to 30 seconds or less.

messages = [
    {"role": "system", "content": "You are Sunflower, a helpful assistant made by Sunbird AI who knows many African languages."},
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio": audio_array}, # 16kHz mono numpy array
            {"type": "text", "text": "Please transcribe this Luganda audio."},
        ],
    },
]

vLLM (Text-Only Inference)

vLLM is highly recommended for fast deployments. Because this is a multimodal model, text-only vLLM setup requires disabling vision/audio inputs to save memory.

import os
os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"

from vllm import LLM, SamplingParams
from transformers import AutoProcessor

MODEL_ID = "Sunbird/Sunflower-Gemma4-E2B"

llm = LLM(
    MODEL_ID,
    dtype="bfloat16",
    max_model_len=1024,
    limit_mm_per_prompt={"image": 0, "audio": 0}, # Text-only mode
    enforce_eager=True,
)
processor = AutoProcessor.from_pretrained(MODEL_ID)

def make_prompt(text, target_lang):
    messages = [
        {"role": "system", "content": "You are Sunflower, a helpful assistant made by Sunbird AI who knows many African languages."},
        {"role": "user", "content": f"Translate from English to {target_lang}: {text}"}
    ]
    return processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

sampling = SamplingParams(temperature=0.0, max_tokens=256)

examples = [
    ("Good morning, how are you?", "Luganda"),
    ("The children are going to school today.", "Acholi"),
    ("I would like to buy three kilograms of rice.", "Swahili"),
]
outputs = llm.generate([make_prompt(t, l) for t, l in examples], sampling)
for (text, lang), out in zip(examples, outputs):
    print(f"[{lang}] {text}\n     -> {out.outputs[0].text.strip()}\n")

Transformers (Multimodal / Audio)

import torch
import librosa
import numpy as np
from transformers import AutoModelForMultimodalLM, AutoProcessor

MODEL_ID = "Sunbird/Sunflower-Gemma4-E2B"
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype=dtype, device_map="auto", attn_implementation="sdpa"
).eval()

# Load Audio
audio, _ = librosa.load("audio.wav", sr=16_000, mono=True)
audio = np.asarray(audio, dtype=np.float32)
if len(audio) > 480_000:
    raise ValueError("Please provide no more than 30 seconds of audio.")

messages = [
    {"role": "system", "content": "You are Sunflower, a helpful assistant made by Sunbird AI who knows many African languages."},
    {"role": "user", "content": [
        {"type": "audio", "audio": audio},
        {"type": "text", "text": "Please transcribe this Luganda audio."}
    ]}
]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(
    text=[prompt], audio=[audio], return_tensors="pt",
    text_kwargs={"padding": False, "truncation": False, "add_special_tokens": False},
    audio_kwargs={"max_length": 480_000}
).to(model.device)
inputs["input_features"] = inputs["input_features"].to(dtype=dtype)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=256, do_sample=False)

new_tokens = output[0][inputs["input_ids"].shape[-1]:]
raw_response = processor.decode(new_tokens, skip_special_tokens=False)
print(processor.parse_response(raw_response)["content"])

Training

The model was adapted via the following stages:

1. Continued pretraining on multilingual African-language text (documents, parallel text, transcripts, web, and synthetic text).
2. Speech SFT on audio-transcription examples.
3. Mixed text-and-audio SFT on translation, instruction, and speech-transcription tasks.
4. GRPO post-training to improve translation and general response quality using task-specific reward signals.

Evaluation

Text Translation

Translation quality measured as chrF on the Sunbird/salt69 test split (100 examples per language).

Speech Transcription

Reported ASR results span 110 dataset configurations from Sunbird/speech. Performance varies substantially by language and recording source.

(Rates shown on a 0–1 scale; lower is better. WER can exceed 1.0 when insertions outnumber reference words.)

Limitations

• Preview weights: Output will change before the stable release.
• Uneven multilingual quality: Lower-resource languages have substantially lower automatic scores. The low-resource end of the table is genuinely weak, and output there should be verified by a speaker before use.
• Translation variants: BLEU and chrF compare against a single reference and can penalize acceptable wording, dialect, or orthographic variants.
• Speech variation: Accent, dialect, code-switching, background noise, and domain vocabulary can materially affect transcription.
• Language prompting: The model was trained with language-specific transcription prompts and may perform poorly when the spoken language is omitted or named incorrectly.
• Hallucinations: Like any LLM, it can produce fluent, confident, but incorrect translations or transcripts. This matters more in low-resource languages where errors are harder for non-speakers to spot.

License

Released under the Apache 2.0 license, matching the Gemma 4 base model.

Citation

@misc{sunbird2026sunflowergemma4e2b,
  title        = {Sunflower-Gemma4-E2B},
  author       = {{Sunbird AI}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Sunbird/Sunflower-Gemma4-E2B}},
  note         = {Preview release: Multilingual text and speech model for African languages}
}