Models & Labs
NeoMME: Efficient Multimodal and Multilingual Encoder Released
Hugging Face Bloghigh confidence
Why it matters
- →NeoMME integrates text and image processing in a single model, reducing complexity and computational overhead.
- →It offers significant improvements in visual document retrieval efficiency and storage requirements.
- →The model's architecture supports dynamic image resolution and long bidirectional context, enhancing adaptability.
Hugging Face has introduced NeoMME, a new family of multilingual multimodal encoders designed to process text and images using a single Transformer. This approach eliminates the need for separate pretrained vision towers, streamlining the model's architecture. NeoMME is available in two sizes, 260M and 800M, and is optimized for visual document retrieval, offering significant improvements in efficiency and storage. The model's release marks a step forward in integrating multimodal processing within a unified framework, providing developers with a powerful tool for various applications.
Read originalMore from Hugging Face Blog
© Hugging Face BlogResearchresearch
Fine-tuning 350M Model for Structured Outputs
Hugging Face has demonstrated how fine-tuning a 350M model can significantly enhance its ability to produce structured outputs, a crucial task for many real-world applications. By using a targeted fine-tuning approach with a LoRA adapter and specific reward functions, the model's performance on the IFStruct benchmark improved, achieving a 22.6% pass rate. This approach shows that smaller models can be optimized to match the performance of larger models in specific tasks, making them more viable for integration into downstream systems. The process is accessible, with the fine-tuning runnable on a free-tier GPU, making it a practical option for developers looking to enhance model performance without extensive resources.
Hugging Face Blog
© Hugging Face BlogVideo & Creative AIimage
AI Model Trained to Paint Watercolours
A language model has been trained to paint watercolours using JavaScript and the p5.brush library, capturing the attention of over 1.5 million viewers. The project uses reinforcement learning to teach the model aesthetic preferences, allowing it to create unique, handmade-style paintings. This approach contrasts with the typical perfection of AI-generated images, offering a fresh perspective on AI's creative potential. The open-source nature of the project invites further exploration and replication by others interested in the intersection of AI and art. The model's ability to generate code that paints images reflects a deeper level of interaction between AI and artistic expression. By limiting the model to ten methods from the library, the project emphasizes creativity within constraints. This initiative not only highlights AI's evolving role in art but also encourages a broader dialogue about the future of AI-driven creativity.
Hugging Face Blog
© Hugging Face BlogCoding Toolscoding
Hugging Face Introduces Funes for Agent Memory
Hugging Face has launched Funes, a tool that provides a durable memory layer for coding agents like Claude Code and Codex. This innovation allows agents to index, retrieve, and rank session logs, enabling them to recall past decisions and rationales during their workflow. Funes operates locally, ensuring privacy and control, and can also sync with a private Hugging Face dataset for broader accessibility. This development enhances the functionality of coding agents by allowing them to seamlessly integrate past knowledge into current tasks, making them more efficient and context-aware.
Hugging Face Blog
More in Models & Labs
Models & Labsmodels
llama.cpp b10794 Release Expands Platform Support
The latest b10794 release of llama.cpp continues its trend of broadening platform compatibility, now supporting a wide array of systems including macOS, Linux, Windows, and openEuler. Notably, this update includes support for Vulkan and ROCm 10.0 on Ubuntu, as well as CUDA 13 on Windows, which enhances performance options for developers using these platforms. While the release doesn't introduce new model architectures, it solidifies llama.cpp's position as a versatile inference runtime across diverse hardware configurations. This update is a testament to llama.cpp's commitment to accessibility and performance optimization for developers working with AI models.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10795 release enhances SYCL fusion
The b10795 release of llama.cpp brings notable improvements in SYCL fusion, specifically by combining operations like RMS_NORM+MUL+ADD and ADD+ADD. This enhancement, under GGML_SYCL_ENABLE_FUSION, boosts performance for supported data types, while unsupported combinations revert to standard methods. The update continues to support a wide array of platforms, including macOS with Apple Silicon, Ubuntu with Vulkan, and Windows with CUDA 12 and 13. This makes llama.cpp a robust choice for developers working across different hardware environments, ensuring efficient AI processing and broad compatibility.
llama.cpp Releases
Models & Labsmodels
llama.cpp b10796 Release Adds n_expert_used_max Function
The latest release of llama.cpp, b10796, introduces the n_expert_used_max function, enhancing the model's ability to handle expert layers. This update addresses previous issues where models with expert layers failed to load due to missing checks. By implementing this function, the software can now better manage the number of experts per layer, ensuring smoother model loading and operation. This release doesn't introduce new models but focuses on refining the existing infrastructure to support more complex configurations.
llama.cpp Releases