Slop TVNewsLatest
News

Maya 2 Native opens streaming speech in 12 languages

Maya Research published Apache 2.0 weights and inference code on October 25, with 20 voices and cloning from 3 to 10 seconds of speech.

By

·2 min read

Illustration: Maya 2 Native opens streaming speech in 12 languages
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Maya Research released the weights and inference code for Maya 2 Native on October 25, 2026, under Apache 2.0.
  • Maya 2 Native covers 12 languages (11 Indian plus English), runs 2.50B unique parameters, and streams 24 kHz mono audio in 80 ms frames.
  • The project reports one H100 NVL serving 224 concurrent streams, and says an RTX 4090 with 24 GB fits 48, while a single seat needs about 9.2 GB at load.
  • The model and code are Apache 2.0, but the Kyutai Mimi codec Maya 2 Native depends on stays CC-BY-4.0.

Maya Research published the open weights for Maya 2 Native on October 25, 2026, a streaming text-to-speech model that covers 11 Indian languages plus English. The model card lists the checkpoint as release 1.0 and puts the weights and the inference code under Apache 2.0.

The weights sit on Hugging Face, and the inference engine is on GitHub. The release ships a Python API, an OpenAI-compatible speech endpoint and WebSocket streaming, so a developer can host the model instead of renting Maya Research's hosted API.

Maya 2 Native holds 2.50 billion unique parameters, about 5.0 GB in BF16, and generates 24 kHz mono audio that streams in 80 ms frames. It carries 20 built-in voices and does zero-shot voice cloning from 3 to 10 seconds of reference speech plus its exact transcript. The reference language can differ from the output: the model card says a Hindi reference can become Tamil speech.

The architecture is a Qwen3-1.7B backbone with a depth decoder of about 555 million parameters that fills in the remaining codebooks, decoded to audio by Kyutai's Mimi codec.

Running it needs an NVIDIA GPU. The documented single-seat loading peak is about 9.2 GB plus CUDA context, and the project's memory table scales from 7.7 GB steady at one seat to 15.8 GB at 32. The project reports one H100 NVL serving 224 concurrent WebSocket streams at 323 seconds of audio per second, and an RTX 4090 with 24 GB fitting 48 streams. Those are Maya Research's own measurements, not independent tests.

The license is split. The model and code are Apache 2.0, but Mimi stays CC-BY-4.0, and the card asks redistributors to keep that attribution.

Open weights are not the same as easy access. A write-up on DEV Community made the point directly: 5 GB of BF16 weights means little to a developer without a GPU, so the practical path for most people is still the hosted endpoint. Maya Research raised about $1.9 million in seed money led by South Park Commons, Indian Startup News reported on June 15, 2026.

Download the weights from Hugging Face and the engine from GitHub, or call the hosted API if you would rather not run a GPU.

Sources

  1. huggingface.co - model card: release date, languages, parameters, license, cloning, architecture
  2. github.com - inference engine: memory table, H100 NVL and RTX 4090 concurrency figures
  3. dev.to - co-founder's release post (disclosed); concurrency and latency table
  4. analyticsindiamag.com - independent write-up: 12 languages, 80 ms frames
  5. llms.blog - independent write-up: 2.50B parameters, depth decoder, license split
  6. ayu-sh-kr.com - independent write-up: 9.2 GB load, serving results
  7. dev.to - the "not open in practice without a GPU" catch
  8. indianstartupnews.com - the $1.9M seed round, led by South Park Commons