Hugging Face

Perplexity AI open-sourced a Rust reimplementation of their Unigram tokenizer, achieving 5x lower latency than Hugging Face's tokenizers crate and reducing CPU utilization by 5-6x in production. The optimizations include double-array trie, bitmap packing, and huge pages.

Artificial Analysis and IBM launch ITBench-AA, a benchmark for agentic enterprise IT tasks focusing on Site Reliability Engineering. Frontier models score below 50%, with Claude Opus 4.7 leading at 47%. The benchmark evaluates models on Kubernetes incident response, requiring diagnosis from logs and traces.

This article details how to deploy a fully local voice conversation pipeline for the Reachy Mini robot, eliminating the need for cloud servers or API keys. It uses a cascaded approach combining VAD, STT, LLM, and TTS, with recommended defaults: llama.cpp with Gemma 4, Silero VAD, Parakeet-TDT 0.6B v3 STT, and Qwen3-TTS. Various LLM options are provided, including local MLX, Transformers, vLLM, or remote Responses API.

This article clarifies often-confused AI agent terms like 'harness' (execution layer) and 'scaffold' (behavior-defining layer), explaining model, agent, tool use, sub-agents, and training concepts.

NVIDIA introduces Nemotron-Labs Diffusion language models that achieve up to 6.4x faster inference than autoregressive models while maintaining high accuracy by generating tokens in parallel and refining them iteratively. The models support three modes: autoregressive, diffusion, and self-speculation. The 8B model outperforms Qwen3 8B by 1.2% accuracy.

A 3-billion-parameter specialized model outperformed all commercial frontier APIs on quality, cost, and stability in an enterprise OCR benchmark, at roughly fifty times lower cost. The result challenges the default assumption that larger models are always better, highlighting distributional alignment as a more decisive performance variable than parameter count.

The open-source movement is bringing AI breakthroughs to robotics, lowering barriers to entry. From the ROS framework to models from Nvidia, Hugging Face, and Alibaba, robots' ability to reason, decide, and act is becoming accessible to more people. However, tensions between commercial incentives and academic ideals present new challenges.

Allen AI released OlmoEarth v1.1, reducing compute costs up to 3x while maintaining v1 performance by merging tokens across resolutions. The new model family is designed for large-scale remote sensing analysis and is already deployed globally by partners.

Six new Sentence Transformers CrossEncoder rerankers, state-of-the-art at their respective sizes, built on Ettin ModernBERT encoders with a distillation recipe. The smallest (17M) to largest (1B) all outperform prior models on MTEB and NanoBEIR. Full training recipe, dataset, and models are open-sourced.

This article presents a parameter-efficient fine-tuning approach using LoRA and DoRA to adapt NVIDIA Cosmos Predict 2.5 for robot video generation on a single GPU. It covers data preparation, adapter initialization, training with rectified flow loss, inference, and evaluation metrics.

IBM Research launches the Open Agent Leaderboard, an open benchmark for comparing full agent systems (not just models). It evaluates generality across six diverse benchmarks, reporting both quality and cost. Early results show general agents are competitive with specialized ones, and agent architecture is increasingly impactful. All code, data, and paper are open-source.

IBM releases Granite Embedding Multilingual R2, two multilingual embedding models (97M and 311M parameters) built on ModernBERT, supporting 32K token context, covering 200+ languages, and achieving top scores on MTEB multilingual retrieval benchmarks. The 97M model is the best sub-100M open model, and the 311M model ranks #2 among models under 500M parameters.

A new open-source model GLiNER2-PII with 0.3B parameters achieves state-of-the-art performance on PII detection, surpassing OpenAI's Privacy Filter on the SPY benchmark. It recognizes 42 entity types and is trained on a multilingual synthetic corpus. The model is publicly available on Hugging Face.

Understanding modern AI architectures is harder than ever. This article introduces a simple trick: replace 'huggingface.co' with 'hfviewer.com' in any Hugging Face model URL to instantly get an interactive visualization of the model's architecture. The tool, hfviewer, supports transformers, vision, and multimodal models with zero setup. Terminal commands and a browser extension offer even faster access.

This article explains how to separate CPU and GPU workloads to achieve a massive performance boost for inference. Continuous batching improves GPU utilization by tightly packing batches, but synchronous operation causes idle gaps where CPU and GPU wait for each other, accounting for nearly a quarter of total runtime. By using non-default CUDA streams and events for asynchronous batching, the CPU and GPU can work in parallel, eliminating idle time and providing a free 24% speedup. The article details CUDA streams, events, and their application to continuous batching, with reference to the implementation in the transformers library.

A malicious Hugging Face repository imitating an OpenAI release delivered infostealer malware to Windows machines, amassing around 244,000 downloads before removal. Researchers warn that public AI model registries pose supply chain risks as developers clone models into corporate environments.

This article examines AWS's infrastructure components for foundation model pre-training, post-training, and inference, including GPU instances, Elastic Fabric Adapter (EFA), Lustre file system, and UltraCluster/UltraServer architectures, emphasizing the role of open-source software in resource management and monitoring.

Unsloth, an open-source AI optimization library, has been officially welcomed into the PyTorch Ecosystem Landscape. The announcement highlights Unsloth's contributions to model training, quantization, and community impact, including over 250M model downloads and being the 10th most-followed organization on Hugging Face. Unsloth will continue its open-source work with closer collaboration with PyTorch.

MachinaCheck is a multi-agent AI system built on AMD MI300X that generates CNC manufacturability reports from STEP files in 30 seconds. It runs fully on-premise to protect intellectual property, combining geometric parsing with LLM reasoning.

Crusoe and NVIDIA Dynamo developed fastokens, an open-source Rust BPE tokenizer that achieves a 9.1× average speedup over HuggingFace and up to 40% faster TTFT on long-context workloads.

In conversation with Clem Delangue, co-founder and CEO at Hugging Face about the state of Open Source AI (and the myth of Sisyphus)

OncoAgent is an open-source, privacy-preserving clinical decision support system for oncology. It features a dual-tier LLM architecture (9B speed vs 27B deep reasoning), multi-agent LangGraph topology, Corrective RAG pipeline over 70+ NCCN and ESMO guidelines, and a three-layer reflexion safety validator with Zero-PHI policy. The system routes queries via complexity scoring and was fine-tuned on AMD Instinct MI300X, achieving 56x throughput acceleration. It supports on-premises deployment to ensure data sovereignty.

CyberSecQwen-4B is a small specialized model fine-tuned from Qwen3-4B-Instruct for defensive cybersecurity, addressing data privacy, cost, and offline deployment needs. It matches or surpasses the 8B Cisco Foundation-Sec-Instruct model on CTI-Bench benchmarks while running on a single consumer GPU. The article details training methodology, data sources, benchmark results, and future directions.

Allen AI releases EMO, a mixture-of-experts model pretrained end-to-end so that modular structure emerges directly from data without human-defined priors. EMO allows using only 12.5% of experts per task while retaining near full-model performance, and works as a strong general-purpose model when all experts are used. Unlike standard MoE, EMO's expert subsets degrade only slightly when selectively used.

gnucleus-ai has released an open-source FreeCAD dataset on Hugging Face, containing 100 parametric CAD models (shafts, bearings, flanges, etc.) with key parameters, images, and FCStd files, suitable for CAD generation tasks. The dataset is licensed under Apache-2.0, includes various mechanical parts, and supports 3D, image, and text modalities.

A detailed walkthrough of LoRA fine-tuning Qwen3-1.7B on the MedMCQA dataset using AMD MI300X with ROCm. The entire pipeline runs without CUDA, training takes ~5 minutes, and the model outputs both answer and explanation.

Learn how to deploy any HuggingFace model in one session using Goose and Together's Dedicated Container Inference. Skip the setup complexity — one prompt gets your model running in a production-grade GPU environment on release day.

ServiceNow AI team's migration from vLLM V0 to V1 for RL training identified four backend fixes: processed logprobs, runtime defaults, inflight weight updates, and fp32 lm_head. They prioritized backend correctness before applying objective-side corrections, achieving full parity with V0 reference.

This article reviews ML Intern, an open-source ML assistant that goes beyond AutoML by supporting the entire workflow from dataset research to model deployment. It demonstrates a practical project: building a text classification model for customer support tickets, covering steps like dataset selection, smoke testing, and training plan approval.

The article explores the rising cost of AI evaluation, especially for agent benchmarks, showing that evaluation has become a new compute bottleneck. Static benchmarks can be compressed 100-200x, but agent and training-in-the-loop benchmarks resist compression. Reliability demands repeated runs, multiplying costs. High evaluation costs risk concentrating validation power in well-funded labs.

DeepInfra joins Hugging Face Hub as an Inference Provider, offering cost-effective serverless inference on over 100 models, starting with conversational and text-generation tasks, accessible via UI and SDKs.

NVIDIA announces Nemotron 3 Nano Omni, a new omni-modal understanding model that processes text, images, video, and audio. Built on a hybrid Mamba-Transformer-MoE backbone with C-RADIOv4-H vision encoder and Parakeet-TDT-0.6B-v2 audio encoder, it achieves top benchmarks in document understanding, ASR, video understanding, and efficiency. Designed for real-world document analysis, automatic speech recognition, long audio-video understanding, agentic computer use, and general multimodal reasoning.

This article explains how to use Scikit-LLM's text summarization feature to handle large volumes of text in machine learning pipelines. It covers building a custom Hugging Face summarizer transformer, integrating it into a scikit-learn pipeline with TF-IDF vectorization and a classifier, and demonstrates the process with code examples.

This article demonstrates building three scalable web applications using OpenAI's newly released open-source Privacy Filter: Document Privacy Explorer, Image Anonymizer, and SmartRedact Paste. Each app showcases different capabilities of the model and leverages gradio.Server for efficient backend processing and custom frontends.

DeepSeek released V4 with a 1M-token context window, optimized for long-running agentic workloads. Its hybrid attention mechanism (CSA and HCA) reduces KV cache to 2% of traditional GQA. Post-training additions include interleaved thinking across tool calls, a dedicated tool-call schema (|DSML|), and DSec sandbox for RL rollouts. The model achieves competitive agent benchmarks.

QIMMA (Arabic for 'summit') is a quality-first Arabic LLM leaderboard that validates benchmarks before evaluation, revealing systematic quality issues in widely-used Arabic benchmarks. It consolidates 109 subsets from 14 benchmarks across 7 domains, applies multi-model automated assessment and human review, and ranks models with a focus on native Arabic capability. The leaderboard is the first to include code evaluation for Arabic LLMs.

This article examines the role of AI in cybersecurity, particularly how the Mythos model uses system-level capabilities to find and patch vulnerabilities. It emphasizes the structural advantage of openness for defense, advocates for semi-autonomous AI agents with human oversight, and notes that open ecosystems are better positioned to counter evolving attacks than proprietary systems.

Ecom-RLVE extends the RLVE framework from single-turn reasoning to multi-turn, tool-augmented e-commerce conversations, providing 8 verifiable environments (product discovery, substitution, cart building, returns, order tracking, policy QA, bundle planning, multi-intent journeys) with procedural problem generation, a 12-axis difficulty curriculum, and algorithmically verifiable rewards. A Qwen 3 8B model trained with DAPO over 300 steps shows early results that environment scaling and adaptive difficulty transfer to agentic, real-world task completion.

DeepSeek releases V2.5-1210 as the final version of the V2.5 series, introducing internet search, improved benchmarks in math, coding, writing, and roleplay, and open-sourcing the model on Hugging Face. The team thanks users and hints at next-gen foundation models.

DeepSeek officially launched DeepSeek-V2.5, merging DeepSeek-V2-0628's general conversational abilities with DeepSeek-Coder-V2-0724's robust code processing. The model shows significant improvements in writing, instruction-following, and safety alignment, and is now available via web, API, and open-source on HuggingFace.