dhee-nxtgen-qwen3-marathi-v1

Dhee-Nxtgen-Qwen3-Marathi-v1

Model Description

Dheegpt-Qwen3-Marathi is a large language model designed for high-quality natural language understanding and generation in Marathi. It is based on the Qwen3 architecture and optimized for both dialogue and reasoning tasks.

The model supports fluent conversational responses and reasoning-style outputs, making it suitable for applications like chatbots, virtual assistants, and step-by-step question answering.

---

Key Features

• Fluent and context-aware Marathi text generation
• Optimized for assistant-style conversations
• Handles summarization, question answering, and open-ended text generation
• Fully compatible with Hugging Face transformers
• Supports integration with VLLM for high-performance batched inference

---

Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "dheeyantra/dhee-nxtgen-qwen3-marathi-v1"

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

# Regular conversation
prompt = "नमस्कार! आजचे हवामान कसे आहे?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Intended Uses & Limitations

Intended Uses

• Chatbots and conversational agents in Marathi
• Story generation, narratives, and creative text
• Domain-specific natural language generation tasks in Marathi

Limitations

• Model outputs are based on the patterns and information in training data; it may generate incorrect or biased information.
• Performance may vary depending on input complexity.
• Primarily designed for Marathi; code-mixed or other languages may yield lower quality outputs.
• May produce irrelevant or hallucinated content occasionally.

---

VLLM / High-Performance Serving Requirements

To serve this model using vLLM, ensure the following:

• GPU with compute capability ≥ 8.0 (e.g., NVIDIA A100).

• PyTorch 2.1+ with CUDA toolkit installed.

• For Tesla V100 (sm70), vLLM GPU inference is not supported; CPU-only fallback is possible but slow.

• Python dependencies:

  pip install torch transformers vllm sentencepiece
  

Example vLLM command:

vllm serve \
  --model dheeyantra/dhee-nxtgen-qwen3-marathi-v1 \
  --host 0.0.0.0 \
  --port 8000

---

License

Released under the Apache 2.0 License.