Open weights, datasets and benchmarks for Indian languages.

Every artifact here is built native-first, verified by people who speak the language, and released in the open. Nothing held back for a paid tier.

3Live releases

2Languages

3Artifact types

100%Open licensed

Release index

Numbered in order of release

Models, datasets and benchmarks, each with weights or data, an evaluation card, and a technical report.

OXFI‑001

Indic TTS

Text-to-speech for Hindi and Marathi. Streaming-ready checkpoints, multiple speakers, trained on native recordings.

Model

hi।mr
MIT।15 Aug 2026

Weights (hi)Weights (mr)

OXFI‑002

ATX IndicEval

One language, one region, full spectrum. A pure evaluation benchmark built from verified native sources, not repurposed training data.

Benchmark

mr।hi
CC BY‑SA 4.0।15 Aug 2026

Dataset Report

OXFI‑003

Romanized Marathi SFT

Code-mixed instruction data at real scale, in the Latin-script Marathi people actually type. Deduplicated, filtered, human-audited.

Dataset

mr‑Latn
CC BY‑SA 4.0।15 Aug 2026

Download

OXFI‑004

IndicEval, next language‑region

Same template, same depth, a new language as it is actually lived. Contributor call opens after the first release.

Upcoming

To be announced

OXFI‑005

Reserved

Slot held for the next model release.

Upcoming

To be announced

Technical reports

Method, data, limitations

ATX IndicEval-Benchmark

A language–region unified evaluation benchmark for Indic LLMs.

Benchmark।4 min

Everything on Hugging Face

Weights · data · eval cards

Every release is mirrored on the ATX Labs organisation, with model cards, data cards and reproducible eval configs alongside the files.

huggingface.co/atx-labs ↗

Use and cite

Free for research and production

Every release ships with a model or data card, an evaluation card, and a reproducible harness. Use it commercially, fine-tune it, redistribute it. Attribution is all we ask.

Models MIT
Benchmarks CC BY-SA 4.0
Datasets CC BY-SA 4.0
Code MIT

@misc{openxfi2026,
  title  = {OpenXFI: Shaping India's AI, Together},
  author = {ATX Labs},
  year   = {2026},
  url    = {https://openxfi.com}
}