Maya 2 Native

Maya Research's 2.5B streaming text-to-speech model for 11 Indian languages and English, with 20 voices and cloning, under Apache 2.0.

In short. Maya 2 Native reads text aloud in Hindi, Tamil, Telugu, Bengali and eight other languages, including English. It starts playing audio while it is still generating, offers 20 built-in voices, and can copy a voice from a short recording. The weights are free under Apache 2.0 and need an NVIDIA graphics card.

Summary

Maya 2 Native is an open-weight streaming text-to-speech model from Maya Research for 11 Indian languages and English: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. It produces 24 kHz speech in 80 ms audio frames, has 20 built-in voices, and clones a voice from a 3 to 10 second reference, including across languages. The model has 2.5B parameters: a backbone initialized from Qwen3-1.7B-Base predicts the first of 32 codebooks of Kyutai's Mimi codec, and a depth decoder of about 555M parameters fills in the rest. Training used licensed and synthetic speech, followed by a preference-optimization stage, and Maya does not disclose its datasets or total training hours.

On its own bench6k set of 6,000 held-out texts, 500 per language, Maya reports an overall character error rate of 3.49% with its in-house Indic speech recognizer as the main judge, 3.40% with Qwen3-ASR and 2.86% with a second in-house recognizer held out from training. Assamese is the hardest language, at 5.41% to 6.59%, Hindi is among the easiest at 1.52% to 2.05%, and web addresses reach 29% on the main judge. On one H100 NVL it served 224 concurrent streams with a median time to first audio of 111 ms, and a single stream took about 14 ms. The card also says Maya-2-Native ranked second for Hindi on VoiceArena's real-time leaderboard on 3 November 2026, behind Cartesia's Sonic-3.6, and seventh in the all-models Hindi view. These are Maya's release measurements from automated judges, and the card reports no human naturalness study and no controlled comparison with other speech models.

The weights and the inference engine use Apache 2.0, while the upstream parts keep their licenses: Qwen3-1.7B-Base is Apache 2.0, Kyutai's Mimi codec is CC BY 4.0 and SpeechBrain's ECAPA-TDNN is Apache 2.0. The engine targets Linux, Python 3.10 to 3.12 and NVIDIA CUDA 12.8, and with a BF16 backbone one seat uses 7.7 GB of GPU memory in steady state and peaks at 9.2 GB while loading. The card does not establish CPU, Apple or browser inference. It offers a Python API, an OpenAI-compatible speech endpoint and WebSocket streaming. In the 30 days to 11 November 2026, Hugging Face counted about 5,600 downloads.

Good for

  • Voice agents, phone menus and accessibility tools for Indian languages that must stream audio from your own GPU
  • Cloning a voice you have permission to use and speaking it in another Indian language
  • Narration and notifications in many Indian languages from one 5 GB model

Watch out for

  • Assamese is the weakest language and some outputs drift toward a Bengali-like timbre, and URLs, symbols, dates and long digit strings stay hard even with text normalization
  • The scores come from automated recognizers and a speaker-embedding classifier run by Maya, and the card reports no human listening study
  • It needs an NVIDIA GPU with CUDA on Linux, and the card does not establish CPU, Apple or browser inference
  • There is no audio watermark, so clone only voices you may use, and the Mimi codec is CC BY 4.0, so keep its attribution
  • Maya does not disclose the training datasets or the total hours of audio

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Maya 2 Native. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Maya-2-Native BF16 2.5B 5 GB 5.75 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Maya 2 Native?
Maya 2 Native reads text aloud in Hindi, Tamil, Telugu, Bengali and eight other languages, including English. It starts playing audio while it is still generating, offers 20 built-in voices, and can copy a voice from a short recording. The weights are free under Apache 2.0 and need an NVIDIA graphics card.
Can I use Maya 2 Native commercially?
Yes. Maya 2 Native is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Maya 2 Native open source?
Maya 2 Native is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Maya 2 Native on my own computer?
Yes. The smallest variant, Maya-2-Native, needs about 5.75 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Maya 2 Native released?
Maya 2 Native was released by Maya Research on October 25, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
OmniVoice Xiaomi 613M – CC BY-NC (Non-commercial) 3.75 GB Laptop or phone
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
VoxCPM2 OpenBMB 2.29B – Apache 2.0 (Permissive) 8 GB Laptop or phone
ZONOS3-Mini Zyphra 800M 100M active – Apache 2.0 (Permissive) 10 GB Laptop or phone