AWS adds Alibaba's Qwen3-TTS voice-cloning model to SageMaker JumpStart
AWS published a guide to deploying Alibaba's publicly available Qwen3-TTS-12Hz-1.7B-Base on SageMaker JumpStart as a managed real-time endpoint for voice cloning.
What happened
AWS published a technical walkthrough for deploying Qwen3-TTS-12Hz-1.7B-Base, a text-to-speech model from Alibaba Cloud's Qwen team, from Amazon SageMaker JumpStart to a fully managed real-time inference endpoint, per the AWS Machine Learning Blog.
The deployment uses the SageMaker Python SDK, with Amazon CloudWatch metrics for monitoring and right-sizing the endpoint; AWS says it handles provisioning, health monitoring, and automatic scaling so users don't manage GPU servers.
The Base variant clones a voice from a short reference clip plus its transcript, applying the speaker's timbre, pitch, and cadence to new text without retraining. This differs from the CustomVoice variant, which draws on a fixed set of predefined speakers.
AWS lists Qwen3-TTS-12Hz-1.7B-CustomVoice and Qwen3-ASR-1.7B as also available in JumpStart.
AWS attributes the following to the model family: 10 supported languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian), streaming generation for low latency, cross-lingual cloning that preserves vocal identity across languages, few-second reference audio, and suitability as a fine-tuning base.
AWS's stated benefits are cost aligned to compute rather than per-character API pricing and data kept within the customer's AWS account and endpoint. These are vendor claims from a deployment tutorial; no independent benchmarks or third-party evaluations are supplied.
AWS names media, education, customer experience, and localization as target uses, and makes no mention of safeguards against voice-cloning misuse.
Anthropic's @ClaudeDevs says Claude Code will now find a graceful stopping point when the 5-hour limit hits mid-task, drawing on a small fixed allowance from the weekly limit.
AWS Machine Learning Blog · Primary source ·