Dhee-NxtGen-Qwen3-Bengali-v2

Model Description

Dhee-NxtGen-Qwen3-Bengali-v2 is a large language model designed for natural and fluent Bengali (Bangla) understanding and generation.
Built upon the Qwen3 architecture, this model is optimized for assistant-style dialogue, function-calling tasks, and reasoning-oriented responses.

It is part of DheeYantra’s multilingual initiative in collaboration with NxtGen Cloud Technologies Private Limited, focusing on building domain-adapted Indic LLMs.

Key Features

Fluent, context-aware Bengali text generation
Fine-tuned for assistant-style interactions and reasoning tasks
Handles open-domain question answering, summarization, and dialogue
Fully compatible with 🤗 Hugging Face Transformers
Optimized for VLLM serving for high-performance inference

Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "dheeyantra/dhee-nxtgen-qwen3-bengali-v2"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# Qwen3-compatible formatted prompt
prompt = """<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
তুমি কি আমার জন্য একটি অ্যাপয়েন্টমেন্ট নির্ধারণ করতে পারবে?<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Intended Uses & Limitations

Intended Uses

Bengali conversational chatbots and assistants
Function-calling and structured response generation
Story generation and summarization in Bengali
Natural dialogue systems for Indic AI applications

Limitations

May generate inaccurate or biased responses in rare cases
Performance can vary on out-of-domain or code-mixed inputs
Primarily optimized for Bengali; other languages may produce less fluent results

VLLM / High-Performance Serving Requirements

For high-throughput serving with vLLM, ensure the following environment:

GPU with compute capability ≥ 8.0 (e.g., NVIDIA A100)
PyTorch 2.1+ and CUDA toolkit installed
For V100 GPUs (sm70), vLLM GPU inference is not supported; CPU fallback is possible but slower.

Install dependencies:

pip install torch transformers vllm sentencepiece

Run vLLM server:

vllm serve   --model dheeyantra/dhee-nxtgen-qwen3-bengali-v2   --host 0.0.0.0   --port 8000

License

Released under the Apache 2.0 License.

Developed by DheeYantra in collaboration with NxtGen Cloud Technologies Pvt. Ltd.

dheeyantra/dhee-nxtgen-qwen3-bengali-v2