Maya 2 Native ships Apache-2.0 streaming TTS for 12 languages

Voice-model watchers on X spent Tuesday pushing Maya Research's Maya 2 Native release across their timelines: a 2.5B-parameter streaming text-to-speech model, weights and code under Apache-2.0, covering 12 languages. The verified fact underneath the buzz: the model card is live, and everything above checks out on it.

Key facts

• The model card confirms 12 languages (11 Indic plus English), 20 built-in voices, zero-shot cloning from 3-10 seconds of reference audio, and cross-lingual synthesis — a Hindi reference voice can produce Tamil speech.
• Audio is 24 kHz mono, streamed in 80 ms frames, with continuous batching and CUDA graphs for concurrent generation.
• Single-seat loading peaks around 9.2 GB of VRAM, and the quickstart recommends a 12 GB or larger NVIDIA GPU with CUDA 12.8.
• The release ships a Python API, an OpenAI-compatible /v1/audio/speech endpoint, and WebSocket streaming, plus the inference code repo on GitHub under Apache-2.0.
• The announcement thread drew ~1.4k likes and 47k+ views in about six hours per the velocity check on October 27 — reach claims, not model claims.

What's driving the conversation

@Dheemanthredy (Dheemanth Reddy BS, on the Maya Research team per the model card) posted the launch thread, with @mayaresearch_ai's account and amplification from @ai_for_success and @noctus91 carrying it through the Indic-AI builder crowd. The tenor of the replies splits pretty cleanly: people building voice agents for Indian languages are excited about the licensing (Apache-2.0 means no per-call fees and no seat-counting), and the TTS-quality crowd is busy comparing it to Fish Audio's Drama 3 and the commercial stacks. Nothing in the discourse disputes the release itself — the argument is about where it lands in the rankings.

The substance

Verified: the Hugging Face model card and the GitHub inference repo are live, both under Apache-2.0, with the checkpoint release stamped 1.0 on October 25, 2026. The card documents the full spec: 2.50B unique parameters (about 5.0 GB of BF16 weights), 24 kHz audio in 80 ms streaming frames, 20 voices, zero-shot and cross-lingual cloning. Upstream components are Qwen3-1.7B-Base for the backbone, Kyutai's Mimi codec, and SpeechBrain's ECAPA-TDNN for speaker conditioning.

Claims to keep stamped: Maya's positioning as the "#1 Hindi TTS" in launch chatter overstates its own evidence. The model card's actual citation is a #2 rank for Hindi among real-time TTS models on VoiceArena — Elo 1047, behind Cartesia's Sonic-3.6 at 1094 — checked October 27, 2026 on a leaderboard that updates as samples arrive. That is a good result, not a dominance claim. And there is one sharp edge for self-hosters: the documented local server ships without built-in authentication, so anyone exposing it needs to put a reverse proxy with auth in front or keep it bound to localhost.

Why builders are watching

Two reasons this one moves. First, the OpenAI-compatible /v1/audio/speech endpoint means most voice-agent stacks can swap it in with a config change — the same integration path our explainer on the OpenAI-compatible API covers. Second, it extends a licensing pattern we have tracked all month, from Fish Audio's Drama 3 to Morena and Vambo's open-weights drop: speech models that were rent-only or license-restricted keep going fully open, and Indic-language coverage — a market of roughly 1.5 billion speakers that commercial TTS has served thinly — is where the open releases are landing hardest. For anyone running self-hosted voice agents in Hindi, Tamil, Bengali, or the other eight languages, the build-versus-rent math just changed in one direction worth recalculating.