Sunflower-Qwen3.8-27B

Sunbird AI's open model for 68 African languages, with the deepest coverage of the 31 languages of Uganda. It is the strongest open-weight model available for translation between English and African languages, it matches Gemini 3.1 Pro when translating into African languages while beating GPT-5.1 and Claude Sonnet 4.5 by a wide margin, and it reasons adaptively: with thinking enabled it decides per prompt whether a reasoning trace is worth it, so maths and knowledge questions get one and translation stays fast.

Model	Base	Use it for
Sunflower-Qwen3.8-27B (this model)	Qwen3.8-27B	Best quality across translation, comprehension and reasoning
Sunflower-Qwen3.5-9B	Qwen3.5-9B	Translation on smaller GPUs, within 0.4 chrF of the 27B

Highlights

• Translation, 68 African languages: chrF 46.2, against 32.3 for the best other open model, 38.9 for Claude Sonnet 4.5 and 38.0 for GPT-5.1.
• Into African languages, the hard direction, Sunflower is level with Gemini 3.1 Pro across the 68-language set (41.6 vs 41.8) and ahead of it by 4.4 chrF on the 31 Ugandan languages (39.2 vs 34.8), producing roughly one thirtieth of the output tokens per sentence.
• FLORES-200: best open model on the public benchmark (chrF 53.5 vs 49.1), ahead of GPT-5.1.
• SAHARA: 42.15, first among open-weight models, sixth overall behind five closed models.
• Comprehension and reasoning: best open model on four of six AfroBench-Lite tasks. With thinking on, AfriMGSM rises from 40.3 to 60.5 while translation quality is unchanged.
• The 9B stands on its own: chrF 45.8 on the 68 languages, ahead of every other open model, including ones three times its size.

Evaluation

Scores are chrF or accuracy on a 0 to 100 scale, thinking off unless stated. Every number comes from the public evaluation code at github.com/SunbirdAI/sunflower, with the same sentences, prompts and output budgets for every model. Sunflower's own starting checkpoints are shown alongside, so the gain from training can be read directly.

Translation

Mean chrF over both directions.

Model	68 African languages¹	31 Ugandan languages²	FLORES-200, 14 languages³
Gemini 3.1 Pro⁴	48.5	42.1	58.1
Sunflower-Qwen3.8-27B	46.2	42.4	53.5
Sunflower-Qwen3.5-9B	45.8	42.0	52.0
Claude Sonnet 4.5	38.9	30.6	n/a⁵
GPT-5.1	38.0	30.8	51.4
AfriqueQwen3.5-9B Instruct v1	32.3	23.3	49.1
AfriqueQwen3.5-4B Instruct v1	31.1	22.3	47.9
Qwen3.8-27B (starting point of the 27B)	29.7	23.0	44.9
Qwen3.5-27B	29.4	21.9	42.9
Qwen3.5-9B (starting point of the 9B)	25.7	20.1	38.0

By direction. Sunflower's lead is largest when translating into African languages, the harder direction and the one most users need.

Model	68 languages, eng→xx	68 languages, xx→eng	Ugandan, eng→xx	Ugandan, xx→eng
Gemini 3.1 Pro⁴	41.8	55.2	34.8	49.4
Sunflower-Qwen3.8-27B	41.6	50.8	39.2	45.7
Sunflower-Qwen3.5-9B	42.1	49.6	39.5	44.5
Claude Sonnet 4.5	35.3	42.5	27.1	34.1
GPT-5.1	33.3	42.7	26.5	35.0
AfriqueQwen3.5-9B Instruct v1	28.8	35.8	22.4	24.3
AfriqueQwen3.5-4B Instruct v1	28.2	34.1	21.4	23.2

¹ Sunbird/salt-69 test set: 68 African languages plus French, 100 sentences per language. ² Sunflower translation eval, test split: 99 sentences per language. ³ FLORES-200 devtest for the 14 AfroBench-Lite languages, 5-shot, 100 sentences per direction. ⁴ Gemini 3.1 Pro runs with reasoning on, which cannot be disabled, at roughly 30 times Sunflower's output tokens per sentence. ⁵ Withheld: a contamination probe found Claude Sonnet 4.5 reproducing the public FLORES English references verbatim. Its eng→xx score, 49.0, is valid and just below Sunflower's 49.3.

Closed models were run through their APIs with a neutral translator system prompt; Sunflower used its own. All models received the same instruction, greedy decoding and a 100-token output budget.

Translation by language

Quality tracks how much training data exists for each language. Tiers use Sunflower-27B's mean chrF on the 68-language set.

• Good (50 and above), 30 languages: Afrikaans, Zulu, Chichewa, Swahili, Xhosa, Somali, Nigerian Pidgin, Lingala, Sotho, Malagasy, Igbo, Amharic, Hausa, Yoruba, Kinyarwanda, Tswana, Luganda, Shona, Ndebele, Kirundi, Runyoro, Ewe, Luo, Lusoga, Acholi, Oromo, Runyankole, Kikuyu, Rukiga, Akan.
• Moderate (35 to 50), 19 languages: Rutooro, Bemba, Ruruuli, Bambara, Lango, Lugungu, Lugwere, Ateso, Kumam, Lugbara, Wolof, Lumasaba, Kabyle, Lunyole, Jopadhola, Bari, Lubwisi, Samia, Rukonjo.
• Basic (below 35), 19 languages: Alur, Karamojong, Dagbani, Kakwa, Berber, Ma'di, Fulani, Kwamba, Aringa, Luhya, Lendu, Dinka, Dagaare, Kupsabiny, Pokot, Ik, Kalenjin, Kanuri, Ikposo. Have a speaker check output in these languages before relying on it.

French is supported as a high-resource reference language.

Scores for all 69 languages (chrF, Sunflower-Qwen3.8-27B)

Language	Code	eng→xx	xx→eng	Mean
Afrikaans	afr	82.9	84.6	83.8
Zulu	zul	68.4	79.7	74.0
Chichewa	nya	69.4	73.9	71.7
French	fra	68.7	73.0	70.9
Swahili	swa	68.1	73.2	70.7
Xhosa	xho	65.1	75.5	70.3
Somali	som	59.7	76.8	68.2
Nigerian Pidgin	pcm	57.1	78.4	67.8
Lingala	lin	60.8	71.7	66.2
Sotho	sot	61.4	70.9	66.1
Malagasy	mlg	61.1	70.6	65.9
Igbo	ibo	61.1	70.2	65.6
Amharic	amh	50.3	77.5	63.9
Hausa	hau	57.7	66.1	61.9
Yoruba	yor	51.5	71.0	61.3
Kinyarwanda	kin	59.6	62.7	61.2
Tswana	tsn	57.2	64.8	61.0
Luganda	lug	58.5	63.3	60.9
Shona	sna	52.5	63.5	58.0
Ndebele	nbl	47.4	68.5	57.9
Kirundi	run	50.6	64.7	57.6
Runyoro	nyo	52.0	61.4	56.7
Ewe	ewe	52.0	60.2	56.1
Luo	luo	52.2	57.2	54.7
Lusoga	xog	46.0	61.7	53.9
Acholi	ach	51.2	55.0	53.1
Oromo	orm	46.5	58.4	52.4
Runyankole	nyn	49.5	55.4	52.4
Kikuyu	kik	40.2	62.6	51.4
Rukiga	cgg	47.2	54.9	51.1
Akan	aka	41.9	58.6	50.2
Rutooro	ttj	43.8	55.9	49.8
Bemba	bem	46.3	52.0	49.2
Ruruuli	ruc	40.4	57.8	49.1
Bambara	bam	39.3	53.9	46.6
Lango	laj	38.8	51.6	45.2
Lugungu	rub	39.2	50.6	44.9
Lugwere	gwr	39.0	50.8	44.9
Ateso	teo	41.1	48.1	44.6
Kumam	kdi	39.3	47.7	43.5
Lugbara	lgg	41.2	45.7	43.4
Wolof	wol	34.4	49.1	41.7
Lumasaba	myx	38.4	44.5	41.4
Kabyle	kab	33.4	49.1	41.2
Lunyole	nuj	36.6	44.3	40.5
Jopadhola	adh	34.9	41.4	38.2
Bari	bfa	34.4	40.8	37.6
Lubwisi	tlj	34.1	41.0	37.5
Samia	lsm	35.8	38.8	37.3
Rukonjo	koo	36.2	37.4	36.8
Alur	alz	29.2	39.0	34.1
Karamojong	kdj	33.2	34.4	33.8
Dagbani	dag	32.4	33.9	33.1
Kakwa	keo	34.3	30.3	32.3
Berber	ber	24.2	38.6	31.4
Ma'di	mhi	25.8	27.9	26.9
Fulani	ful	21.3	32.2	26.8
Kwamba	rwm	26.9	25.9	26.4
Aringa	luc	22.8	29.1	26.0
Luhya	luy	17.7	26.7	22.2
Lendu	led	18.4	24.5	21.4
Dinka	din	17.4	25.5	21.4
Dagaare	dga	16.0	25.4	20.7
Kupsabiny	kpz	17.2	22.7	19.9
Pokot	pok	17.1	20.3	18.7
Ik	ikx	15.7	20.9	18.3
Kalenjin	kln	13.0	22.6	17.8
Kanuri	kau	9.8	22.2	16.0
Ikposo	kpo	4.0	18.8	11.4

Comprehension, understanding and reasoning

AfroBench-Lite, 15 languages (Belebele 12), 5-shot, 100 documents per task. Accuracy, or exact match for AfriMGSM. Best score per task in bold.

Task	Sunflower-27B	Sunflower-9B	AfriqueQwen3.5-9B Instruct v1	AfriqueQwen3.5-4B Instruct v1	Qwen3.8-27B	Qwen3.5-27B	Qwen3.5-9B
Reasoning and comprehension
AfriMMLU	62.8	47.8	54.5	44.4	58.9	62.9	42.1
AfriMGSM	40.3	16.7	24.8	34.9	32.2	35.7	21.3
Belebele	82.2	68.2	69.2	62.4	72.0	75.2	58.8
Language understanding
AfriXNLI	69.9	61.5	71.6	68.1	66.9	66.3	57.8
SIB-200	86.5	83.5	84.7	77.5	82.4	82.8	72.4
Injongo intent	87.7	81.9	80.5	67.2	83.7	85.1	71.2

Sunflower-27B leads four of six tasks and is within 0.1 of the best on AfriMMLU. The 9B is tuned for translation; for comprehension and maths, use the 27B.

Adaptive thinking. With thinking enabled the model reasons where it helps and answers directly elsewhere, so maths and knowledge improve and everything else holds (generative AfroBench-Lite tasks, Sunflower-27B):

Task	Thinking off	Thinking on
AfriMGSM	40.3	60.5
AfriMMLU	66.6	71.6
Belebele	84.6	84.3
SIB-200	88.0	87.3
Translation (chrF, 3 languages)	64.3	64.4

SAHARA

Scores from the SAHARA evaluation server, 18 tasks in four clusters. Sunflower-27B is the top open-weight model on the board and sixth overall, behind five closed models.

Cluster	Sunflower-27B, adaptive	Sunflower-27B, thinking off	Sunflower-9B, adaptive	Qwen3.8-27B	Qwen3.5-27B
Knowledge, comprehension and reasoning	64.8	56.5	43.5	69.4	54.6
Text classification	45.2	45.2	39.5	42.4	43.6
Text generation	17.7	17.7	16.1	11.3	11.7
Token-level tasks	40.9	41.4	13.6	45.4	33.8
SAHARA score	42.15	40.22	28.19	42.12	35.95

Training for African languages lifts text generation by 6.4 points and classification by 2.8 over the starting checkpoint, with adaptive thinking closing most of the knowledge gap.

Usage

The model follows the Qwen3.8 chat template. Leave thinking enabled (the default) for general use; the model decides per prompt whether to reason. For bulk translation, set enable_thinking=False: translation quality is the same and generation is faster. Use greedy decoding for translation and a temperature around 0.7 for open-ended generation.

Prompting

Translation:

Translate to {language}: {text}

{language} is the English name of the target language (Luganda, Acholi, Swahili). The source language is inferred; naming it (Translate from Luganda to English: ...) helps with short or ambiguous input.

For other tasks, write the instruction in English and state the reply language (... Reply in Yoruba.), or write the instruction in the target language: the model answers in the language it is addressed in. Multi-turn chat works as normal.

vLLM

import os

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

from transformers import AutoProcessor
from vllm import LLM, SamplingParams

MODEL_ID = "Sunbird/Sunflower-Qwen3.8-27B"

llm = LLM(
    MODEL_ID,
    quantization="fp8",  # ~55 GB of weights -> ~28 GB
    dtype="bfloat16",
    max_model_len=2048,
    gpu_memory_utilization=0.9,
    limit_mm_per_prompt={"image": 0, "video": 0},  # text only
)
tokenizer = AutoProcessor.from_pretrained(MODEL_ID).tokenizer


def make_prompt(text: str, target_language: str) -> str:
    """Build a translation prompt with thinking disabled."""
    return tokenizer.apply_chat_template(
        [{"role": "user", "content": f"Translate to {target_language}: {text}"}],
        add_generation_prompt=True,
        enable_thinking=False,
        tokenize=False,
    )


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(text, lang) for text, lang in examples],
    SamplingParams(temperature=0.0, max_tokens=256),
)
for (text, lang), out in zip(examples, outputs):
    print(f"[{lang}] {text}\n  -> {out.outputs[0].text.strip()}")

Transformers

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

MODEL_ID = "Sunbird/Sunflower-Qwen3.8-27B"

model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID, dtype=torch.bfloat16, device_map="auto"
).eval()
tokenizer = AutoProcessor.from_pretrained(MODEL_ID).tokenizer

messages = [
    {"role": "user", "content": "Translate to Luganda: The children are going to school today."}
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
).to(model.device)
output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(
    tokenizer.decode(
        output[0][inputs["input_ids"].shape[1] :], skip_special_tokens=True
    ).strip()
)

With thinking enabled the reply starts with a <think>...</think> block; allow max_new_tokens of 2048 or more for reasoning tasks.

Training

Starting from Qwen3.8-27B: continued pretraining on African-language text; supervised fine-tuning for translation, comprehension, tagging and instruction following; then reinforcement learning (GRPO) with rewards for translation quality, factuality, language consistency and instruction following. Adaptive reasoning was trained so the model chooses when to think.

Limitations

• Coverage is uneven. Quality tracks training data; see the tiers above, and check basic-tier output with a speaker before use.
• Translation is the most optimised path. General chat and reasoning work well, but the starting checkpoint keeps an edge on general knowledge and NER (see SAHARA).
• Fluent is not the same as correct. Like any LLM it can produce confident, wrong output, and errors are harder to spot in low-resource languages.

License

Apache 2.0.

Citation

@misc{sunflower_qwen38_27b,
  title        = {Sunflower-Qwen3.8-27B},
  author       = {Sunbird AI},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Sunbird/Sunflower-Qwen3.8-27B}}
}