Models & Labs
NVIDIA Launches Cosmos 3 for Physical AI
Hugging Face Bloghigh confidence
Why it matters
- →Cosmos 3 unifies multiple AI capabilities into a single model, simplifying development.
- →It enhances the ability to simulate and understand complex physical environments.
- →Integration with Hugging Face Diffusers facilitates easy adoption and use.
NVIDIA has unveiled Cosmos 3, an omni-model designed for physical AI applications, integrating multiple capabilities into a single framework. This model, built on a Mixture-of-Transformers architecture, allows for seamless world generation, scene understanding, and policy generation. Cosmos 3 supports various modalities, including text, image, video, and action, enabling developers to simulate and understand complex physical environments. Available on Hugging Face, it offers two versions: Cosmos 3 Nano for efficient inference and Cosmos 3 Super for large-scale data generation. This release aims to streamline the development of AI systems in robotics, autonomous vehicles, and smart spaces.
Read originalMore from Hugging Face Blog
© Hugging Face BlogModels & Labsmodels
JetBrains Releases Mellum2: Efficient 12B MoE Model
JetBrains has unveiled Mellum2, a 12 billion parameter Mixture-of-Experts model designed for efficient text and code processing. By activating only 2.5 billion parameters per token, Mellum2 offers more than twice the inference speed of similar-sized models, making it ideal for high-throughput, latency-sensitive tasks. This model is particularly suited for software engineering applications, such as code generation and summarization, and can be deployed in private environments due to its open-source Apache 2.0 license. Mellum2 represents a shift towards specialized, efficient models that enhance the performance of larger AI systems without replacing them.
Hugging Face Blog
© Hugging Face BlogResearchagents
Scalable Enterprise AI Hinges on Agent Logic
Hugging Face's exploration into agent logic reveals its potential to transform enterprise AI adoption. By integrating agent logic, which includes software primitives like knowledge graphs and algorithms, AI agents can more effectively navigate complex enterprise workflows. This approach reduces token consumption and enhances performance, as demonstrated in IBM's use of agents for tasks like legacy code understanding and test generation. The shift towards agentic AI could lead to more cost-effective and reliable AI solutions in enterprise settings, marking a significant step forward in scalable AI deployment.
Hugging Face Blog
More in Models & Labs
Models & Labsmodels
llama.cpp adds EXAONE 4.5 implementations
The latest llama.cpp release expands its capabilities with the integration of EXAONE 4.5, bringing new vision markers and projector paths into the fold. This update aligns EXAONE 4.5 with the Qwen2.5-VL-style encode path, enhancing model loading and tensor registration processes. Developers will find improved performance and compatibility, particularly when working with EXAONE models. While no new models are introduced, the release refines existing functionalities, ensuring robust performance across various systems. This step forward is crucial for developers seeking to leverage EXAONE 4.5's full potential.
llama.cpp Releases
Models & Labsmodels
llama.cpp b9455 Release Adds Quantized KV Cache
The latest b9455 release of llama.cpp introduces quantized KV cache support, a notable enhancement for efficiency in AI model inference. This update also addresses a partial view fix and removes an overly strict assert, improving the overall robustness of the software. While the release includes various platform builds, the focus remains on optimizing performance across different environments. The addition of quantized KV cache support is a step forward in making AI models more resource-efficient, particularly beneficial for developers working with limited computational resources.
llama.cpp Releases
Models & Labsmodels
llama.cpp b9457 release focuses on Vulkan improvements
The latest b9457 release of llama.cpp brings a notable improvement in Vulkan performance by reducing host memory lock contention, which can enhance efficiency in certain workloads. This update replaces unique_lock with lock_guard, aiming to streamline operations. While the release doesn't introduce new models or major features, it continues to refine the platform's compatibility across various systems, including macOS, Linux, and Windows. The focus remains on optimizing existing capabilities rather than expanding into new territories.
llama.cpp Releases