Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

NVIDIA opens weights for Nemotron 3.5 Transcribe

AI summary

NVIDIA opened the model weights for Nemotron 3.5 Transcribe, an open-weight MoE speech model with about 34 billion total parameters and 4.3 billion active per forward pass, according to TMTPost. The model combines transcription, speaker attribution, translation, word-level timestamps and audio question answering, supporting 40 transcription languages and as many as eight speakers in one hour of audio under the reported OpenMDW-1.1 license.

In-depth insight

The release gives developers a locally runnable multilingual speech model that combines transcription with attribution, translation and audio Q&A, expanding the open-weight tooling available for agent and edge applications.

From the digest “Edge AI Daily 早报(10月31日)”

十一、英伟达开源语音转录模型,8 人抢话也能分清

1.产品:英伟达于2026年10月28日正式开放Nemotron 3.5 Transcribe模型权重,采用MoE架构,总参数约340亿、每次前向仅激活约43亿;将转录、说话人归属、翻译、单词级时间戳乃至音频问答整合进单一检查点,支持40种语言转录、31种语言与英语互译、单次处理1小时录音并区分最多8名说话人,以OpenMDW-1.1许可开放,单卡即可本地运行。

2.数据与趋势:在Hugging Face Open ASR英语组平均词错误率5.33%,居开放权重模型首位,Artificial Analysis测试下约40亿激活参数对应AA-WER 2.8%,位于准确率与激活参数的帕累托前沿;配套开放Granary v2数据清洗流水线,覆盖463,062小时、48种语言(较上一代25种近乎翻倍),意在用更低推理成本与数据可控性撬动闭源语音API的按量计费模式。

3.观点:该模型隶属Nemotron 3.5开放家族(同期另有可本地运行的300亿参数Nemotron 3.5 Lightning),定位为智能体的“耳朵”,配合微调工具、数据流水线把部署推向边缘与本地,最终导向英伟达GPU需求;文章判断开源语音基座时代来临,闭源API护城河收缩至成本与信任,但实时流式转录、极端噪声与超长会议仍待验证,OpenMDW-1.1的商业边界亦待明确。

(广角观察、Edge AI Daily等综合整理)

#NVIDIA#Nemotron 3.5 Transcribe#Granary v2#OpenMDW-1.1#Hugging Face#ModelScope

Original · 1 min readRead the original
Group chat

No comments yet. Start the conversation.