Models & Labs
Nemotron 3.5 Enhances Multimodal AI Safety
Hugging Face Bloghigh confidence
Why it matters
- →Nemotron 3.5 addresses the challenge of multimodal safety by evaluating text, images, and responses together.
- →The model supports custom policy enforcement, making it adaptable to different industry needs.
- →Its reasoning capabilities provide transparency and accountability, essential for compliance in regulated environments.
Hugging Face has unveiled Nemotron 3.5, an advanced AI safety model that evaluates multimodal inputs—text, images, and responses—together to identify policy violations. This model allows for custom policy enforcement, making it suitable for various industries with different safety requirements. It maintains multilingual support and introduces a safety dataset, enhancing its applicability in global markets. Nemotron 3.5's reasoning capabilities provide detailed justifications for its safety verdicts, crucial for compliance and audit purposes. This release represents a significant advancement in AI safety, offering enterprises a more adaptable and accountable content moderation tool.
Read originalMore from Hugging Face Blog
© Hugging Face BlogResearchresearch
EVA-Bench Data 2.0 Expands to 213 Scenarios
EVA-Bench Data 2.0 significantly broadens its scope by expanding from one to three enterprise domains, covering Airline Customer Service Management, Enterprise IT Service Management, and Healthcare HR Service Delivery. This update quadruples the scenario coverage to 213, offering a robust benchmark for evaluating voice agents across diverse workflows. The scenarios are meticulously validated against leading models like OpenAI GPT-5.4 and Google Gemini 3.1 Pro, ensuring they are both challenging and fair. This release not only enhances the realism and variety of the dataset but also sets a new standard for reproducibility and authentication in voice agent evaluation.
Hugging Face Blog
© Hugging Face BlogCoding Toolscoding
Hugging Face CLI Optimized for Coding Agents
Hugging Face has revamped its command-line interface (CLI) to better accommodate both human users and coding agents like Codex and Claude Code. The updated CLI now auto-detects when it's being used by an agent and adjusts its output format accordingly, providing a more efficient and token-light experience. This change significantly reduces the token usage for complex tasks, making it more efficient for agents to interact with the Hugging Face Hub. The CLI's new features include agent-mode output and enhanced logging methods, which streamline multi-step tasks and improve usability for both humans and agents.
Hugging Face Blog
© Hugging Face BlogResearchresearch
DharmaOCR Uses DPO to Reduce Text Degeneration
Hugging Face's DharmaOCR has demonstrated a novel application of Direct Preference Optimization (DPO) to significantly reduce text degeneration in OCR tasks. Unlike traditional supervised fine-tuning, which often fails to address degeneration directly, DPO uses the model's own degenerate outputs as negative training signals. This approach led to an average reduction in degeneration rates by 59.4%, with some cases seeing reductions as high as 87.6%. By focusing on the structural failure modes of models, DharmaOCR offers a new methodology for improving model performance in structured tasks without relying on subjective human judgments.
Hugging Face Blog
More in Models & Labs
Models & Labsmodels
v0.22.1 fixes CUTLASS fmin compatibility
The v0.22.1 release of vLLM addresses a critical compatibility issue with CUTLASS fmin during the initialization of DeepSeek-V4. This update ensures that users relying on this configuration experience smoother integration and improved functionality. By resolving this specific technical challenge, the release contributes to the ongoing refinement and stability of the vLLM framework. Users can now expect enhanced performance and fewer compatibility problems, reinforcing the platform's reliability. This update is a testament to the continuous efforts to maintain and improve the technical robustness of vLLM.
vLLM Releases
Models & Labsmodels
llama.cpp b9509 release optimizes token processing
The b9509 release of llama.cpp brings a key optimization by preventing unnecessary checkpoint restores when new tokens are detected. This update ensures that the system only applies a conservative -1 subtraction when no new tokens are present, thereby minimizing redundant KV state restoration. Developers working with token-based tasks will find this change streamlines processing and boosts efficiency. While the release doesn't introduce new models or architectures, it enhances the runtime's performance across macOS, Linux, and Windows, including support for ROCm 7.2 and CUDA 12 and 13. This makes llama.cpp more efficient and adaptable for developers using different hardware configurations.
llama.cpp Releases
Models & Labsmodels
llama.cpp b9510 release enhances WASM SIMD128 support
The latest b9510 release of llama.cpp introduces significant optimizations for the ggml_vec_dot_q4_1_q8_1 function using WASM SIMD128 intrinsics. This update focuses on improving performance by vectorizing the inner loop, which is crucial for efficient computation in WebAssembly environments. The changes are specifically gated to ensure non-WASM builds remain unaffected, maintaining broad compatibility. This release marks a step forward in optimizing AI model inference on diverse hardware, particularly benefiting those leveraging WebAssembly for AI workloads.
llama.cpp Releases