Meta 发布 Muse Spark Open MoE 模型,可用 NVIDIA NeMo 微调
Meta 这次放的多模态推理模型直接支持音频和视频输入,配合 NVIDIA NeMo AutoModel 不用转换格式就能 LoRA 微调。做垂直领域适配的工程师可以先下 NIM 试跑,再决定要不要做全量 SFT。
Meta 发布了新的 Muse Spark Open MoE 模型,开发者可以通过 NVIDIA NeMo AutoModel 在领域数据集上做后训练。NeMo AutoModel 是 PyTorch 原生训练库,原生支持 Hugging Face checkpoint,无需模型格式转换。
作为推理模型,Muse Spark Open 在图像和文档理解之外增加了音频和视频感知能力。AutoModel 开箱即用地支持全量 SFT 和 LoRA 微调,针对多种行业场景的快速实验做了优化,NVIDIA 还提供了可下载的 NIM 版本。
对开发者来说,免转换的 checkpoint 加载降低了把开源多模态推理模型适配到垂直领域的门槛,NIM 下载也便于本地部署验证。
Developers can post-train @Meta's new Muse Spark Open MoE model on domain specific datasets using NVIDIA NeMo AutoModel, a PyTorch native training library with native @huggingface checkpoint support and no model conversion required.
As a reasoning model, Muse Spark Open adds audio and video perception alongside image and document understanding
AutoModel enables full SFT and LoRA fine tuning out of the box, optimized for rapid experimentation over a wide variety of industry specific cases. Try the downloadable @nvidia NIM.
Read the full post: https://t.co/iZVp4oOH32
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Muse Spark Open, NVIDIA NeMo, Meta, LoRA.
Source attribution: curated from PyTorch (@PyTorch). Cross-source corroboration: 11 independent sources have covered this development (including IT之家, TestingCatalog, Tech in Asia), confirming industry-wide signal strength.