Instructions to use dheeyantra/dhee-pravega with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dheeyantra/dhee-pravega with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="dheeyantra/dhee-pravega")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("dheeyantra/dhee-pravega")
model = AutoModelForMultimodalLM.from_pretrained("dheeyantra/dhee-pravega", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dheeyantra/dhee-pravega with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dheeyantra/dhee-pravega"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dheeyantra/dhee-pravega",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dheeyantra/dhee-pravega

SGLang

How to use dheeyantra/dhee-pravega with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "dheeyantra/dhee-pravega" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dheeyantra/dhee-pravega",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "dheeyantra/dhee-pravega" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dheeyantra/dhee-pravega",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use dheeyantra/dhee-pravega with Docker Model Runner:
```
docker model run hf.co/dheeyantra/dhee-pravega
```

Dhee-Pravega

Dhee-Pravega is a compact, low-latency agentic tool-calling model for Indian languages. Give it your function schema and a user message in any of 13 Indian languages and it replies in that language while emitting the correct tool call -- using Gemma-4's native tool-calling format, so it drops straight into the standard tokenizer.apply_chat_template(messages, tools=...) flow with no bespoke prompting.

Version: beta-01 · Languages: 13 Indian languages — Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Sanskrit, Sindhi, Tamil, Telugu — plus English (typically used for the system prompt and tool schema).

What this model is for

Build in-language voice and chat agents: pass your tools + the user's message, get back a correct <|tool_call> and a reply in the user's own language. Ideal for IVR/voice bots, chat assistants, and on-device agents across Indian languages where English-only tool models fall short.

What you get

⚡ Lightning-fast tool calling — sub-50 ms to first token on Blackwell GPUs, with a verified ~47 ms even on an RTX 4090 (prewarmed, merged bf16, ~100-token tool-calling prompt), so the tool call begins almost the instant the user stops speaking.
🌐 Excellent accuracy across 13 Indian languages — Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Sanskrit, Sindhi, Tamil and Telugu (plus English) — picking the right function and replying in-language in every one.
🗣️ Stays in the user's language — keep your system prompt and tool schema in English; the model still answers the user in their native language, consistently, instead of drifting back to English.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("dheeyantra/dhee-pravega")
model = AutoModelForCausalLM.from_pretrained("dheeyantra/dhee-pravega", torch_dtype=torch.bfloat16,
                                             device_map="auto")

# System prompt carries the tool JSON schema (English); ask in any of 13 Indic
# languages and the model responds in-language, emitting function calls.
messages = [
    {"role": "system", "content": "You are a helpful assistant with access to ..."},
    {"role": "user", "content": "मुझे दिल्ली का मौसम बताओ"},
]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
ids = tok(prompt, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=256)[0][ids.input_ids.shape[1]:]))

A GGUF (Q4_K_M) build for llama.cpp is included when present in this repo.

Evaluation

metric	value
first token latency (ms, this GPU, warm)	27

Limitations

A small (E2B) model -- not a general reasoning or long-context LLM; best on short, tool-oriented exchanges.
Tool-calling fluency is strongest for tools resembling the training distribution (Glaive-style function schemas).
Inherits the Gemma base model's terms and its prohibited-use policy.

Files in this repository

README.md
chat_template.jinja
config.json
generation_config.json
model.safetensors
tokenizer.json
tokenizer_config.json

Licence

Released under GEMMA — free to use, share and adapt for non-commercial purposes with attribution.

Commercial use requires a separate licence — contact contact@dheeyantra.com.

This is a non-commercial release regardless of any upstream permissive label; see the attribution below for the full licence lineage.

Attribution

Base model: google/gemma-4-E2B-it, governed by the Gemma Terms of Use (which continue to apply to this derivative, including the prohibited-use policy).
Data: dheeyantra/function-calling-indic (Glaive-style tool-use conversations translated to 13 Indian languages).

Citation

@misc{dhee_pravega_ckpt-721db047,
  title  = {Dhee-Pravega},
  author = {DheeYantra Research Labs},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/dheeyantra/dhee-pravega}},
  note   = {Non-commercial release; contact contact@dheeyantra.com for commercial use}
}

Downloads last month: -

Safetensors

Model size

5B params

Tensor type

BF16

Model tree for dheeyantra/dhee-pravega

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Finetuned

(293)

this model