Droplychee-1.0-27B

Developed by: droplychee (Md Mushfiqur Rahim) Model type: Text generation (multimodal — image-text-to-text) Parameters: 28B Precision: BF16 License: MIT Languages: English, Bengali Base model: None — trained from scratch

Model Description

Droplychee-1.0-27B is a 28B-parameter language model trained from scratch (no base/foundation model). It supports text generation and image-text-to-text tasks, with support for English and Bengali.

Intended Use

• General-purpose conversational AI
• Bengali and English text generation
• Multimodal image understanding tasks

How to Use

With Transformers

from transformers import pipeline

pipe = pipeline("text-generation", model="droplychee-core/droplychee-1.0-27b")
messages = [
    {"role": "user", "content": "Bangladesh er rajdhani ki?"}
]
pipe(text=messages)

With vLLM

pip install vllm
vllm serve "droplychee-core/droplychee-1.0-27b"

curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "droplychee-core/droplychee-1.0-27b",
    "messages": [
      {"role": "user", "content": "Bangladesh er rajdhani ki?"}
    ]
  }'

Evaluation Results

Benchmarks were run using lm-evaluation-harness.

Benchmark	Score
MMLU (overall)	84.05%
GSM8K (flexible-extract)	81.05%
ARC-Challenge (acc_norm)	65.61%
HellaSwag (acc_norm)	83.89%

MMLU category breakdown

Category	Accuracy
STEM	83.89%
Humanities	78.75%
Social Sciences	90.45%
Other	85.90%

Limitations

Citation

@misc{droplychee2026,
  title  = {Droplychee-1.0-27B},
  author = {Md Mushfiqur Rahim},
  year   = {2026},
  url    = {https://huggingface.co/droplychee-core/droplychee-1.0-27b}
}