What is Maya 2 Native?
Maya 2 Native is an open-weights text-to-speech model from Maya Research, an Indian start-up. Checkpoint 1.0 was published on Hugging Face on 25 October 2026, after a closed preview earlier in the month. It speaks Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, with 20 built-in voices and zero-shot cloning from 3 to 10 seconds of reference audio; a voice cloned in Hindi can speak Tamil.
It has 2.5 billion parameters, about 5.0 GB in BF16: a backbone initialised from Qwen3-1.7B-Base predicts audio tokens, a 555M-parameter decoder fills in the rest, and Kyutai's Mimi codec turns them into 24 kHz mono speech in 80 ms frames that can be streamed as they are made.
Weights and code are Apache 2.0; the Mimi codec it downloads is CC BY 4.0. The package ships a Python API, an OpenAI-compatible speech endpoint and WebSocket streaming. Maya Research publishes no hosted API price.
Is Maya 2 Native worth it?
Maya 2 Native is an open speech model that independent listeners place close to the commercial leaders in Hindi. On VoiceArena's Hindi board it sits seventh at 1047 Elo from 2,268 votes, level with Mistral's Voxtral Mini TTS (1048) and Eleven v3 (1044) and ahead of Sarvam's Bulbul-V3 (957), but 67 points behind Eleven v4 at 1114. The card's "second" applies only with the board filtered to real-time models, where Cartesia's Sonic-3.6 leads at 1094; the published rank interval of fifth to tenth is the fairer reading.
The serving numbers are the strongest part of the release. On one H100 NVL the team ran 224 simultaneous WebSocket streams at 323 seconds of audio per second, with a median 111 ms and 99th-percentile 242 ms to first audio and no stalls. One seat needs 7.7 GB of GPU memory and 32 seats 15.8 GB, so memory is not the constraint on a 24GB card, though no throughput figure is published for anything smaller than an H100.
Outside Hindi the evidence is self-graded. Maya's 6,000-text benchmark puts overall character error rate at 2.9% to 3.5% depending on the judge, and up to 6.6% in Assamese, but two of the three judges are its own speech recognisers, there is no human naturalness study and no cloning-similarity score. The weak spots are listed plainly: Assamese has no voice trained on even an hour of it and drifts towards a Bengali timbre, web addresses score 29% CER, and very short replies lose voice identity.
Deployment needs care. The official package runs only on Linux with an NVIDIA GPU and CUDA 12.8, though community ports to Apple's MLX already exist; the bundled server has no authentication; and the audio carries no watermark. Training data is described only as licensed and synthetic speech, with no hours or sources given.
Where it earns its place
A voice-agent engineer at a Bengaluru lending start-up
Read out EMI reminders and payment confirmations in the borrower's language
Serve the WebSocket endpoint on one GPU with max_seats sized to call volume; the card's 224-stream run fits on one H100 NVL. Check numbers, dates and account digits by ear first, since written forms remain a known failure mode.
A product lead at an edtech company with Hindi and Tamil courses
Narrate lessons in one consistent brand voice across languages
Clone a consenting presenter from a 3 to 10 second clip with its transcript and generate in any of the 12 languages. There are no per-character fees, but clone similarity is approximate and no score is published.
A platform team already using Bolna for phone agents
Add Indian-language speech without building a TTS server
Select Maya 2 Native as Bolna's voice provider; audio is downsampled from 24 kHz to 8 kHz for calls, and language can switch mid-call without reconnecting.
Before you commit
- Setup
- pip install the maya3 package from GitHub on Linux with Python 3.10 to 3.12 and a CUDA 12.8 NVIDIA GPU, then Maya3TTS.from_pretrained; first use downloads the weights, Mimi and the ECAPA speaker encoder.
- Integrations
- Python generate and stream APIs, an OpenAI-compatible /v1/audio/speech endpoint, WebSocket streaming, Bolna's voice-agent platform; community MLX ports for Apple silicon.
- Data and privacy
- Self-hosted, so text and voice references stay on your servers. Apache 2.0 weights; training data described only as licensed and synthetic speech. No audio watermark, and the bundled server has no authentication.
- Cost at scale
- GPU time only: 7.7 GB of GPU memory for one seat and 15.8 GB for 32, and the card's single H100 NVL served 224 concurrent streams at 323 seconds of audio per second.
- Limits
- 24 kHz mono; best effort on emotion tags; weak on Assamese, URLs, symbols and one- or two-word turns; batching changes the random stream, so seeds only reproduce in isolation.
What’s good
- 1047 Elo in Hindi on VoiceArena, level with Eleven v3
- Apache 2.0 weights and server
- 224 streams on one H100 NVL at 111 ms median first audio
- Cross-lingual cloning from 3 to 10 seconds of audio
What’s not
- 67 Elo behind Eleven v4 in Hindi; no independent ranking for other languages
- Most benchmark judges are Maya's own ASR models
- Assamese has no native voice; URLs score 29% CER
- NVIDIA and Linux only officially; no audio watermark
Best for
Developers building voice agents, IVR or narration in Indian languages who want to self-host under a permissive licence.
Maya 2 Native alternatives
Maya 2 Native: common questions
- Is Maya 2 Native worth it?
- For Indian-language voice agents, Maya 2 Native is the open model to start with: close to the paid leaders in Hindi and cheap to run at scale, with automated scores standing in for proof everywhere else. We rate it 4.0 out of 5.
- How much does Maya 2 Native cost?
- Free under Apache 2.0 (Mimi codec CC BY 4.0); self-hosted on your own GPUs. No first-party hosted price published.
- What are the downsides of Maya 2 Native?
- 67 Elo behind Eleven v4 in Hindi; no independent ranking for other languages; most benchmark judges are Maya's own ASR models; assamese has no native voice; URLs score 29% CER; nVIDIA and Linux only officially; no audio watermark.
- Who is Maya 2 Native best for?
- Developers building voice agents, IVR or narration in Indian languages who want to self-host under a permissive licence.
- How much does Maya 2 Native cost for a team or in production?
- GPU time only: 7.7 GB of GPU memory for one seat and 15.8 GB for 32, and the card's single H100 NVL served 224 concurrent streams at 323 seconds of audio per second.
- Is Maya 2 Native safe for company data?
- Self-hosted, so text and voice references stay on your servers. Apache 2.0 weights; training data described only as licensed and synthetic speech. No audio watermark, and the bundled server has no authentication.
- What are the best Maya 2 Native alternatives?
- The closest alternatives we have reviewed are Cartesia Sonic 3.6, Eleven v4 and v4 Turbo, OmniVoice and Qwen3-TTS.
Answers checked Nov 2, 2026.
Make Maya 2 Native? Show this rating on your site
