gnani-evon-v3.3-30B-A3B

gnani-evon-v3.3-30B-A3B

Chats, answers questions and reasons in English and ten Indic languages with a 128K context. A 30B mixture-of-experts model from Gnani.ai, for teams building assistants, RAG and agents in Indian languages.

Made by gnani on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: gnani-evon-v3.3 is a text model from Gnani.ai built for Indian languages. It mixes Mamba and transformer layers in a mixture-of-experts design, with 30B total parameters and about 3.5B active per token. Training had three stages, more pretraining, supervised tuning and then reinforcement learning, and the model accepts inputs up to 128K tokens.

Who it suits: Teams that build chat assistants, document search or agents in Hindi, Tamil, Bengali and other Indic languages. The weights carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Build assistants that answer in Indic languages | Replying to a question written in Telugu script
Knowledge management | Platform engineers | Run multilingual retrieval over long documents | Searching long files in several Indic languages
Software | Agent developers | Build agents that call tools across several turns | An assistant that uses a tool during a Hindi chat

Anyone who needs advice on high-stakes matters can skip it, since the card advises against that use.

Where to use it:

• Transformers 5.3.0 or newer, with trust_remote_code=True
• vLLM 0.12.0 or newer
• SGLang
• Linux machines with NVIDIA H100 80GB, H200 or A100 GPUs, the hardware the card lists
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Indic chat | App builders | Replies in the user's own script
Long-context RAG | Platform teams | Answers grounded in documents up to 128K tokens
Tool-calling agents | Developers | Multi-turn reasoning with tool calls

How to set it up:

• Python, plus the huggingface_hub package
• A Linux machine with one of the GPUs the card lists
• Room for the weights, about 63.5 GB in BF16

1. Install Python and the huggingface_hub package with the first command below.
2. Download the model files with the second command.
3. Install Transformers 5.3.0 or newer, or vLLM 0.12.0 or newer, or SGLang.
4. Load the model with trust_remote_code=True.

pip install -U huggingface_hub
hf download gnani/gnani-evon-v3.3-30B-A3B

How to use it: The card's Transformers example loads the tokenizer and model in BF16 with remote code allowed, then sends one chat message. It uses the shorter id gnani/gnani-evon-v3.3, so swap in the repository id you downloaded if the two differ. The input is chat messages with system, user, assistant and tool roles, and the output is text that can include a reasoning trace.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "gnani/gnani-evon-v3.3"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Explain photosynthesis in Hindi."},
]

Example: You send the message "Explain photosynthesis in Hindi." as the card's snippet does. The model should answer in Hindi text, and it may add a reasoning trace before the answer.

For newcomers and experts: This is a large model for GPU servers, so a first step is to read the card and download the files. Skip the code until you have a Linux machine with one of the listed GPUs.

Underneath, the card calls it a Mamba2-Transformer hybrid mixture of experts, listed as a Nemotron hybrid MoE network in BF16, with a context of 131,072 tokens. Only a small share of the parameters is active for each token, which keeps compute lower than the total size suggests.

Download the files with hf
Load with remote code allowed
Send chat messages
The model reasons and answers
Read the text reply

Good to know:

• License: Apache 2.0
• Cost to run: no API key or paid service named, but the weights need about 63.5 GB and a data-center class GPU
• Data it reaches: the download fetches files from Hugging Face, and the source names nothing else
• Not for: high-stakes decisions, legal, medical or financial advice, or use without safety checks in your own languages
• Languages: English plus ten Indic languages, among them Hindi, Tamil and Odia

Title: MILU macro mean over 11 languages, as the card reports
gnani-evon-v3.3: 78.74
Sarvam-105B: 75.71
Sarvam-30B: 67.15

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 31.7B
• Weights: about 63.5 GB in BF16; about 15.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 60,825
• Likes: 81
• License: apache-2.0
• Languages: en, hi, bn, te, ta, mr, gu, kn
• Quantized versions: 1
• Tags: transformers, safetensors, nemotron_h, text-generation, nemotron, pytorch, mamba, indic, long-context, reasoning, tool-calling, conversational

Links

• Model: https://huggingface.co/gnani/gnani-evon-v3.3-30B-A3B
• Files: https://huggingface.co/gnani/gnani-evon-v3.3-30B-A3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gnani/gnani-evon-v3.3-30B-A3B

Then follow the usage in its model card. Nothing in this kit runs by itself.