Models & Labs
Meta unveils Muse Glimmer: a 30B parameter multimodal model
Hugging Face Bloghigh confidence
Why it matters
- →Muse Glimmer's open-source nature allows for widespread adoption and experimentation by developers.
- →The model's ability to handle both images and videos alongside text expands the potential applications of AI in multimedia contexts.
- →Advanced attention mechanisms improve processing efficiency, making it a practical choice for real-world applications.
Meta has introduced Muse Glimmer, a new 30 billion parameter multimodal model available on the Hugging Face Hub. This model features a 2 billion parameter vision encoder and a 28 billion parameter text decoder, designed to handle both images and videos. It employs advanced attention mechanisms to improve processing efficiency. Muse Glimmer is open source, providing developers with immediate access to integrate it into their projects. This release marks a significant advancement in multimodal AI capabilities.
Read originalMore from Hugging Face Blog
© Hugging Face BlogAgentsagents
Strands Agents Integrate with Hugging Face for Seamless AI Training
Hugging Face has introduced a streamlined process for recording, training, and deploying AI models using Strands Agents, LeRobot, and Hugging Face Storage Buckets. This integration allows developers to continuously collect data, train policies, and deploy them without the need for repeated data transfers. By utilizing Hugging Face's object-storage repository, users can efficiently manage datasets and checkpoints, reducing overhead and improving workflow efficiency. This development simplifies the AI training loop, making it more accessible and cost-effective for developers working with robotic systems.
Hugging Face Blog
© Hugging Face BlogModels & Labsmodels
OlmoEarth Studio Offers Custom Embedding Exports
OlmoEarth Studio has rolled out a feature that lets users generate and export custom embedding vectors from Earth observation data. These embeddings, crafted using OlmoEarth's open-source foundation models, open up possibilities for tasks like similarity search and segmentation. Users can tailor their embeddings by choosing specific parameters such as the area of interest, time span, and encoder variant. This advancement makes it easier and more cost-effective to utilize Earth observation data for diverse applications, allowing researchers and developers to conduct more precise and efficient analyses without needing extensive labeled datasets.
Hugging Face Blog
© Hugging Face BlogModels & Labsmodels
LFM2.5-VL-3B Enhances Vision-Language Capabilities
Hugging Face's LFM2.5-VL-3B model marks a significant step forward in vision-language processing, offering improved capabilities in screen understanding, object grounding, and multi-image reasoning. By integrating a SigLIP2 400M NaFlex vision encoder and doubling its vocabulary, the model excels in multilingual visual comprehension and tool use. It supports on-device inference, making it accessible for high-volume workloads with impressive speed and efficiency. This release positions LFM2.5-VL-3B as a versatile tool for developers needing robust vision-language solutions that can operate efficiently on a range of devices.
Hugging Face Blog
More in Models & Labs
Models & Labsmodels
llama.cpp b10412 Release Enhances Backend Sampling
The latest b10412 release of llama.cpp introduces backend sampling for both dflash and dspark, marking a technical enhancement in the platform's capabilities. This update allows for more refined control with the enablement of p_min > 0 in backend sampling, adding a layer of precision for developers. While the release doesn't introduce new models or architectures, it quietly strengthens the platform's backend functionality, making it more versatile for developers working across various systems. This update is a step forward in optimizing the performance and flexibility of llama.cpp's inference capabilities.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10414 Release Adds TQ2_0 Support
The b10414 release of llama.cpp marks a significant enhancement with the addition of GGML_TYPE_TQ2_0 type processing in the Metal backend, enabling ternary operations with 2 bits per element. This update brings a more efficient mul_mv kernel, focusing on float operations and optimizing data handling through techniques like precalculating sums. While the release doesn't feature new models, it refines the platform's performance and broadens its compatibility across systems like macOS, Linux, and Windows. By improving efficiency and versatility, llama.cpp continues to be a valuable tool for developers working with a variety of hardware configurations.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10418 Release Enhances SYCL Support
The b10418 release of llama.cpp brings notable improvements to SYCL support, particularly through the introduction of host pinned memory, which enhances host-to-device memory access. This update also resolves a thread-safety issue, ensuring more stable performance across different hardware setups. While no new models are introduced, the release focuses on strengthening the existing infrastructure, making it more robust for developers working with SYCL. This update is crucial for optimizing performance and ensuring compatibility, especially for those leveraging SYCL in their development environments.
llama.cpp Releases