Research
4-bit Model Outperforms Full-Precision Original
Hugging Face Bloghigh confidence
Why it matters
- →QAH allows compressed models to outperform their full-precision counterparts, challenging traditional assumptions about quantization trade-offs.
- →The method enhances training stability, reducing the risk of model degradation during deployment.
- →It offers significant efficiency gains, enabling high-performance models to run on smaller hardware.
Hugging Face has unveiled a new method called Quantization-Aware Healing (QAH) that enables a 4-bit model to outperform its full-precision version. This approach involves distilling knowledge from the original, uncompressed model rather than a recovered checkpoint, allowing the smaller model to achieve superior accuracy and stability. Applied to a GPT-OSS 120B model, QAH resulted in a compressed model that surpassed its full-precision counterpart on most benchmarks. This breakthrough suggests that compressed models can be both efficient and highly accurate, challenging the traditional trade-offs in model quantization.
Read originalMore from Hugging Face Blog
© Hugging Face BlogModels & Labsmodels
Granite 4.2 LLMs Introduced by IBM and Hugging Face
Granite 4.2 marks a significant step forward in reasoning-focused language models, offering three sizes—3B, 8B, and 30B—all built on a dense, decoder-only architecture. These models are pre-trained on a massive 15 trillion tokens and feature a unique five-phase training strategy that extends the context window to 512K tokens. Notably, the 8B and 30B models undergo agentic reinforcement learning, enabling them to operate as agents in real environments, such as editing code and searching the web. This release under the Apache 2.0 license makes advanced reasoning capabilities more accessible to developers, with native tool calling and OpenAI-compatible endpoints enhancing usability.
Hugging Face Blog
© Hugging Face BlogCoding Toolscoding
Gradio Introduces AI Workflow Automation
Gradio's new feature, gr.Workflow, transforms AI workflows by allowing users to create complex pipelines with a simple drag-and-drop interface. This innovation enables users to build, run, and deploy AI applications seamlessly, turning each workflow into a REST API with minimal effort. By integrating with Hugging Face Inference Providers and Gradio Spaces, users can execute multi-step processes and access intermediate results in real-time. This development significantly lowers the barrier for developers to create and deploy sophisticated AI applications, making it easier to experiment and iterate on ideas.
Hugging Face Blog
More in Research
© Google Research BlogResearchagents
AgentHands: Enhancing XR with Interactive Hand Gestures
AgentHands is a groundbreaking prototype from Google Research that integrates expressive hand gestures into XR environments, enhancing the way AI agents interact with users. By synchronizing gestures with speech, AgentHands transforms abstract verbal instructions into intuitive physical demonstrations, making interactions more natural and engaging. This innovation leverages the spatial understanding of XR to provide a more immersive experience, bridging the gap between linguistic intent and physical action. The result is a more human-centric approach to AI, reducing cognitive load and making complex tasks more accessible.
Google Research Blog
© EleutherAI BlogResearchresearch
EleutherAI Explores AI Lie Detection in Aletheia's Quest
EleutherAI's participation in Aletheia's Quest, a competition focused on AI lie detection, has shed light on the intricacies of identifying AI-generated falsehoods. Organized by Cadenza Labs and NDIF, the event tasked teams with developing lie detectors using both black-box and white-box methods on models with up to 120 billion parameters. EleutherAI discovered that black-box monitoring can be surprisingly effective, while white-box probes often struggle outside their training scenarios. This research highlights the challenges in evaluating AI deception as models become more advanced, particularly in detecting subtle forms of deception that go beyond simple factual errors.
EleutherAI Blog
Researchresearch
MIT AI predicts extreme weather without past data
MIT engineers have developed an AI tool that forecasts extreme weather events without relying on historical disaster data. This innovation, called Extreme Event Aware or η-learning, allows for the prediction of statistically-possible events that have not yet occurred, offering new insights for city planners and insurers. By using point statistics and spatial maps, the tool can generate scenarios like a storm with unprecedented rainfall levels. This approach could significantly enhance preparedness for rare but potentially devastating weather events, providing a new layer of resilience planning.
AI News