Instructions to use Sunbird/Sunflower-Gemma4-E2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Sunbird/Sunflower-Gemma4-E2B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Sunbird/Sunflower-Gemma4-E2B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Sunbird/Sunflower-Gemma4-E2B")
model = AutoModelForMultimodalLM.from_pretrained("Sunbird/Sunflower-Gemma4-E2B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Sunbird/Sunflower-Gemma4-E2B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Sunbird/Sunflower-Gemma4-E2B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sunbird/Sunflower-Gemma4-E2B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Sunbird/Sunflower-Gemma4-E2B

SGLang

How to use Sunbird/Sunflower-Gemma4-E2B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Sunbird/Sunflower-Gemma4-E2B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sunbird/Sunflower-Gemma4-E2B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Sunbird/Sunflower-Gemma4-E2B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sunbird/Sunflower-Gemma4-E2B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Sunbird/Sunflower-Gemma4-E2B with Docker Model Runner:
```
docker model run hf.co/Sunbird/Sunflower-Gemma4-E2B
```

Sunflower-Gemma4-E2B

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:

Continued pretraining on multilingual African-language text (documents, parallel text, transcripts, web, and synthetic text).
Speech SFT on audio-transcription examples.
Mixed text-and-audio SFT on translation, instruction, and speech-transcription tasks.
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).

Languages	Directions	Macro BLEU	Macro chrF	Mean Length Ratio
69	138	12.20	0.3413	1.06

Click to view full per-language translation results

Code	Language	EN→X BLEU	EN→X chrF	X→EN BLEU	X→EN chrF	Mean chrF
afr	Afrikaans	57.68	0.7747	57.86	0.7649	0.7698
pcm	Nigerian Pidgin	32.76	0.6438	48.27	0.7306	0.6872
fra	French	34.08	0.6008	40.55	0.6532	0.6270
swa	Swahili	30.77	0.6126	38.06	0.6270	0.6198
zul	Zulu	19.95	0.5949	34.19	0.5887	0.5918
nya	Chichewa	25.38	0.6288	31.98	0.5537	0.5913
xho	Xhosa	15.57	0.5595	29.89	0.5357	0.5476
som	Somali	15.22	0.5180	31.48	0.5647	0.5413
mlg	Malagasy	19.62	0.5458	23.64	0.4907	0.5182
nbl	Ndebele	12.35	0.5109	29.58	0.5239	0.5174
ibo	Igbo	29.19	0.5273	24.54	0.4913	0.5093
lin	Lingala	23.92	0.5414	23.37	0.4670	0.5042
kin	Kinyarwanda	17.54	0.5148	25.29	0.4906	0.5027
sot	Sotho	21.60	0.5077	26.18	0.4950	0.5013
hau	Hausa	20.52	0.5102	24.03	0.4885	0.4993
amh	Amharic	14.52	0.3752	32.33	0.5735	0.4743
lug	Luganda	14.99	0.5089	22.69	0.4384	0.4737
tsn	Tswana	21.32	0.5051	20.01	0.4370	0.4710
sna	Shona	9.52	0.4642	23.02	0.4762	0.4702
run	Kirundi	10.80	0.4388	23.95	0.4554	0.4471
nyo	Runyoro	12.07	0.4644	18.33	0.4224	0.4434
nyn	Runyankole	9.33	0.4415	15.84	0.4114	0.4264
yor	Yoruba	19.98	0.3985	20.58	0.4467	0.4226
cgg	Rukiga	8.30	0.4172	16.60	0.4139	0.4155
xog	Lusoga	7.48	0.4086	15.87	0.3784	0.3935
ttj	Rutooro	8.18	0.3906	14.71	0.3868	0.3887
ewe	Ewe	13.83	0.4096	10.80	0.3265	0.3680
ruc	Ruruuli	5.14	0.3643	15.93	0.3628	0.3636
kik	Kikuyu	6.41	0.3459	14.91	0.3739	0.3599
luo	Luo	9.89	0.3885	12.86	0.3255	0.3570
ach	Acholi	10.98	0.4167	10.99	0.2872	0.3519
orm	Oromo	6.10	0.3643	8.88	0.2961	0.3302
aka	Akan	9.37	0.3244	10.23	0.3186	0.3215
gwr	Lugwere	3.98	0.3349	9.97	0.3074	0.3211
bem	Bemba	5.01	0.3305	10.24	0.3084	0.3195
laj	Lango	6.32	0.3329	9.13	0.2715	0.3022
kdi	Kumam	6.10	0.3176	7.24	0.2739	0.2957
rub	Lugungu	3.35	0.2826	9.40	0.3040	0.2933
nuj	Lunyole	4.38	0.2994	7.73	0.2608	0.2801
koo	Rukonjo	2.72	0.3139	7.14	0.2415	0.2777
tlj	Lubwisi	2.75	0.2743	8.86	0.2780	0.2762
myx	Lumasaba	3.18	0.2784	8.34	0.2674	0.2729
lsm	Samia	3.42	0.2859	9.46	0.2595	0.2727
alz	Alur	3.86	0.2605	5.24	0.2423	0.2514
adh	Jopadhola	4.52	0.2624	6.54	0.2371	0.2498
lgg	Lugbara	4.68	0.2869	4.00	0.2114	0.2491
kab	Kabyle	5.10	0.2313	6.91	0.2438	0.2375
teo	Ateso	3.38	0.2498	5.13	0.2187	0.2343
wol	Wolof	3.99	0.2148	5.60	0.2309	0.2228
bfa	Bari	2.88	0.1759	6.28	0.2441	0.2100
rwm	Kwamba	1.87	0.2103	5.19	0.2068	0.2086
ful	Fulani	3.24	0.1802	6.16	0.2354	0.2078
dag	Dagbani	3.71	0.2026	5.77	0.2106	0.2066
bam	Bambara	5.34	0.1920	4.87	0.2170	0.2045
keo	Kakwa	2.42	0.1812	4.37	0.2086	0.1949
ber	Berber	3.26	0.1749	4.08	0.1973	0.1861
mhi	Ma'di	2.29	0.1575	4.82	0.2127	0.1851
luy	Luhya	0.98	0.1500	5.49	0.2198	0.1849
led	Lendu	1.83	0.1356	3.69	0.2203	0.1779
kdj	Karamojong	1.50	0.1586	3.69	0.1842	0.1714
pok	Pokot	1.19	0.1231	3.60	0.1817	0.1524
ikx	Ik	0.90	0.1060	4.69	0.1970	0.1515
kpz	Kupsabiny	0.72	0.0973	3.68	0.1915	0.1444
dga	Dagaare	1.02	0.0952	3.79	0.1912	0.1432
luc	Aringa	1.07	0.0909	4.49	0.1908	0.1409
kau	Kanuri	0.68	0.0848	4.09	0.1971	0.1409
kln	Kalenjin	0.78	0.1094	4.14	0.1715	0.1405
din	Dinka	1.24	0.0859	2.97	0.1869	0.1364
kpo	Ikposo	0.14	0.0401	1.71	0.1526	0.0964

Speech Transcription

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

Languages	Scored Examples	Macro Language WER	Macro Language CER	Macro Mean Utterance WER	Macro Mean Utterance CER
51	5201	0.4720	0.1890	0.5110	0.2016

Click to view full per-language transcription results

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

Code	Language	Scored	Corpus WER	Corpus CER
ach	Acholi	92	0.4086	0.1758
afr	Afrikaans	170	0.1413	0.0646
aka	Akan	50	0.4182	0.1779
amh	Amharic	250	0.3256	0.1474
bam	Bambara	149	0.6919	0.3454
bem	Bemba	100	0.5718	0.1392
ber	Berber	50	0.5000	0.1924
cgg	Rukiga	49	0.4254	0.1295
dag	Dagbani	100	0.4627	0.1871
dga	Dagaare	50	0.3920	0.1738
eng	English	100	0.1515	0.0819
ewe	Ewe	78	0.5833	0.2704
fra	French	50	0.1029	0.0525
ful	Fulani	149	0.5396	0.2009
hau	Hausa	287	0.2367	0.0844
ibo	Igbo	247	0.5664	0.2802
kab	Kabyle	50	0.7273	0.3696
kau	Kanuri	50	0.6795	0.2133
kik	Kikuyu	50	0.3798	0.1893
kin	Kinyarwanda	100	0.4653	0.1787
kln	Kalenjin	100	0.6164	0.1815
koo	Rukonjo	47	0.7333	0.2119
kpo	Ikposo	50	0.7418	0.3416
led	Lendu	50	0.5656	0.2646
lgg	Lugbara	50	0.3575	0.0990
lin	Lingala	193	0.3886	0.1800
lth	Thur	41	0.4299	0.2430
lug	Luganda	266	0.4999	0.1669
luo	Luo	149	0.3495	0.1096
luy	Luhya	50	0.4335	0.1015
mlg	Malagasy	49	0.3359	0.1561
myx	Lumasaba	45	0.6237	0.1814
nbl	Ndebele	10	0.5043	0.2032
nya	Chichewa	98	0.5949	0.2212
nyn	Runyankole	94	0.4977	0.1613
orm	Oromo	119	0.4602	0.1488
pcm	Nigerian Pidgin	100	0.3856	0.3043
ruc	Ruruuli	38	0.6701	0.1932
rwm	Kwamba	49	1.2251	0.7426
sna	Shona	147	0.3025	0.0951
som	Somali	149	0.4731	0.1898
sot	Sotho	94	0.3946	0.1780
swa	Swahili	99	0.1619	0.0585
teo	Ateso	50	0.4567	0.1516
tsn	Tswana	98	0.2225	0.0826
ttj	Rutooro	49	0.5753	0.2050
wol	Wolof	135	0.5280	0.2602
xho	Xhosa	126	0.4323	0.1247
xog	Lusoga	43	0.4995	0.1378
yor	Yoruba	249	0.5311	0.2042
zul	Zulu	143	0.3090	0.0863

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}
}