{"version":1,"type":"story","url":"https://digestai.news/story/tencent-hy-simulmt-30b-a3b-hugging-face","json":"https://digestai.news/story/tencent-hy-simulmt-30b-a3b-hugging-face.json","markdown":"https://digestai.news/story/tencent-hy-simulmt-30b-a3b-hugging-face.md","slug":"tencent-hy-simulmt-30b-a3b-hugging-face","headline":"Tencent/Hy-SimulMT-30B-A3B · Hugging Face","summary":"The model processes streaming audio and outputs both translated text and speech while cloning the speaker's voice in real time. Tencent reports a StreamLAAL of approximately 3 seconds on the RealSI long-form audio benchmark, the lowest among compared systems, with a GEMBA score exceeding four of five commercial APIs and all open-source models tested. On RealSI-Sentence, Hy-SimulMT scored 84.7, the highest among all listed systems. In a small-scale test with 18 new recordings from September 2026 and six professional interpreters (three CATTI Level 1, three Level 2), the model achieved translation quality on par with experts while often responding faster. The release includes SimulInfer for inference deployment and SimulTrain for continued training. Deployment requires an NVIDIA GPU with CUDA, Python 3.12+, and optionally a configured TTS service for speech output. The model is available on Hugging Face and ModelScope, and can be tried in the Hy Translate App.","keyPoints":["Hy-SimulMT achieves ~3 second StreamLAAL on RealSI benchmark, lowest among compared systems","Model scores 84.7 on RealSI-Sentence, highest among all listed open-source and commercial systems","Open-source release includes SimulInfer inference service and SimulTrain training tools on Hugging Face"],"whyItMatters":"A low-latency open model for simultaneous speech translation with voice cloning could make real-time cross-language communication practical for meetings and live events without proprietary APIs.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Tencent","Tencent Hunyuan","Hugging Face","ModelScope"],"models":["Hy-SimulMT-30B-A3B","Hy-SimulMT-30B-A3B-Interleave"],"people":["Mao Zheng","Bo Lv","Mingrui Sun","Zheng Li","Fangxu Liu","Tao Chen"]},"firstPublishedAt":"2026-10-22T00:00:00Z","updatedAt":"2026-10-22T00:00:00Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"huggingface.co","title":"tencent/Hy-SimulMT-30B-A3B · Hugging Face","url":"https://huggingface.co/tencent/Hy-SimulMT-30B-A3B","publishedAt":"2026-10-22T00:00:00Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xfrgt5/tencenthysimulmt30ba3b_hugging_face/","points":null}],"thread":{"title":"Open Source AI Multilingual Model Race","url":"https://digestai.news/thread/ibm-releases-granite-speech-5-0-2b-for-multilingual-speech-recognition","storyCount":2},"cite":{"text":"Digest AI, \"Tencent/Hy-SimulMT-30B-A3B · Hugging Face\", 22 October 2026, https://digestai.news/story/tencent-hy-simulmt-30b-a3b-hugging-face","publisher":"Digest AI","title":"Tencent/Hy-SimulMT-30B-A3B · Hugging Face","datePublished":"2026-10-22T00:00:00Z","url":"https://digestai.news/story/tencent-hy-simulmt-30b-a3b-hugging-face"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}