Dhee-NxtGen-Qwen3-Indic (4B)

Model Description

Dhee-NxtGen-Qwen3-Indic is a single, unified 4B-parameter multilingual large language model developed by DheeYantra in collaboration with NxtGen Cloud Technologies Pvt. Ltd.

Built on the Qwen3-4B architecture, the model is created to support assistant-style conversations, reasoning, and function-calling–compatible workflows across 14 Indian (Indic) languages within one shared model.

The model is optimized for native-script generation, consistent multilingual behavior, and cross-lingual generalization.

Supported Languages

This single model supports the following Indic languages:

• Hindi (hi)
• Bengali (bn)
• Tamil (ta)
• Telugu (te)
• Malayalam (ml)
• Gujarati (gu)
• Kannada (kn)
• Marathi (mr)
• Odia (or)
• Punjabi (pa)
• Assamese (as)
• Maithili (mai)
• Sanskrit (sa)
• Sindhi (sd)

Best results are achieved when prompts are written entirely in the target language.

---

Key Features

• Single multilingual 4B model (no per-language checkpoints)
• Fluent, native-script text generation across 14 Indic languages
• Optimized for assistant-style and reasoning-based dialogue
• Supports summarization, Q&A, and long-form generation
• Compatible with function-calling style prompting
• Fully compatible with Hugging Face Transformers
• Ready for high-throughput inference using vLLM

---

Example Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# 1. Configuration
model_name = "dheeyantra/dhee-nxtgen-qwen3-indic"
device = "cuda" if torch.cuda.is_available() else "cpu"

# 2. Load Model and Tokenizer
print(f"Loading model: {model_name}...")
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

# 3. Define the Prompt
# Using the ChatML format expected by Qwen-based architectures
prompt = """<|im_start|>system
You are a helpful multilingual assistant.<|im_end|>
<|im_start|>user
क्या आप मेरे लिए एक अपॉइंटमेंट बुक कर सकते हैं? 
अगर हाँ, तो कृपया मुझसे ज़रूरी जानकारी जैसे तारीख, समय और उद्देश्य पूछिए।<|im_end|>
<|im_start|>assistant
"""

# 4. Process and Generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

print("Generating response...")
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

# 5. Decode and Print
# We only want to print the newly generated text (the assistant's reply)
full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
response = full_output.split("assistant")[-1].strip()

print("-" * 30)
print(f"Assistant: {response}")
print("-" * 30)

Function/Tool Calling Example Usage

import torch
import json
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

# --- 1. MODEL SETUP ---
model_name = "dheeyantra/dhee-nxtgen-qwen3-indic"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

# --- 2. TOOLS & SYSTEM PROMPT ---
tools = [{
    "name": "book_appointment",
    "description": "Book an appointment for the user.",
    "parameters": {
        "type": "object",
        "properties": {
            "date": {"type": "string", "description": "Date in YYYY-MM-DD format"},
            "time": {"type": "string", "description": "Time in HH:MM (24h) format"},
            "purpose": {"type": "string", "description": "The medical reason or department"}
        },
        "required": ["date", "time", "purpose"]
    }
}]

# We provide the current date so the model can resolve "tomorrow" or "next Monday"
SYSTEM_PROMPT = f"""You are a helpful AI assistant.
Today's Date: 2026-01-08 (Thursday).

Available Tools:
{json.dumps(tools, indent=2)}

Rules:
1. If details (date, time, purpose) are missing, ask the user in Hindi.
2. If all details are present, output ONLY a <tool_call> JSON.
3. After a tool result is provided, confirm the booking to the user in Hindi."""

# --- 3. BACKEND FUNCTION ---
def execute_booking(date, time, purpose):
    # Simulated backend logic
    if "10:00" in time: # Simulate a busy slot
        return {"status": "error", "message": "यह समय पहले से बुक है। कृपया कोई और समय चुनें।"}
    return {"status": "success", "id": "APP-9921", "doctor": "Dr. Verma"}

# --- 4. THE INTERACTION ENGINE ---
def run_conversation(user_input, history):
    # Add user input to history
    history.append({"role": "user", "content": user_input})
    
    # Construct ChatML prompt
    prompt = f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
    for msg in history:
        prompt += f"<|im_start|>{msg['role']}\n{msg['content']}<|im_end|>\n"
    prompt += "<|im_start|>assistant\n"

    # Generate
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True).split("assistant")[-1].strip()

    # Check if the model wants to call a tool
    tool_match = re.search(r"<tool_call>(.*?)</tool_call>", response, re.DOTALL)
    
    if tool_match:
        print(f"\n[MODEL REQUESTED TOOL]: {response}")
        call_data = json.loads(tool_match.group(1).strip())
        
        # Execute the function
        result = execute_booking(**call_data['arguments'])
        print(f"[TOOL RESULT]: {result}")
        
        # Feed result back to model for final response
        history.append({"role": "assistant", "content": response})
        history.append({"role": "system", "content": f"Tool Result: {json.dumps(result)}"})
        
        # Final confirmation generation
        return run_conversation("Please confirm the result to the user.", history)
    
    return response

# --- 5. TEST SCENARIO ---
chat_history = []

print("--- Chatbot Started (Today is 2026-01-08) ---")

# Step 1: User provides partial info
query_1 = "मेरे लिए डेंटिस्ट का अपॉइंटमेंट बुक करें।"
print(f"\nUser: {query_1}")
res_1 = run_conversation(query_1, chat_history)
chat_history.append({"role": "assistant", "content": res_1})
print(f"Assistant: {res_1}")

# Step 2: User provides the rest
query_2 = "कल दोपहर 2 बजे।"
print(f"\nUser: {query_2}")
res_2 = run_conversation(query_2, chat_history)
print(f"Assistant: {res_2}")

---

Prompting Guidelines

• Use pure native-language prompts for best fluency
• Avoid heavy code-mixing (e.g., Hinglish-heavy inputs)
• Include a system prompt to stabilize multilingual behavior
• Ask explicitly for step-by-step reasoning when required

---

Intended Uses & Limitations

Intended Uses

• Multilingual Indic chatbots and AI assistants
• Education, governance, and public-sector AI applications
• Content generation and summarization in Indian languages
• Cross-lingual conversational and reasoning systems

Limitations

• May occasionally hallucinate or produce inaccurate facts
• Performance may vary slightly across languages
• Not intended for medical, legal, or safety-critical use cases
• Code-mixed inputs may reduce output quality

---

vLLM / High-Performance Serving

Requirements

• NVIDIA GPU with compute capability ≥ 8.0 (A100 / H100 recommended)
• PyTorch 2.1+ with CUDA installed
• V100 (sm70) GPUs are not supported for vLLM GPU inference

Installation

pip install torch transformers vllm sentencepiece

Run vLLM Server

vllm serve   --model dheeyantra/dhee-nxtgen-qwen3-indic   --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.