Loading…
Hugging Face
AI platform and community providing machine learning models, datasets, libraries, and tools for building AI applications.
- 317 posts summarised
- latest
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
From their blog
Latest articles
VibeGame: Exploring Vibe Coding Games
Vibe coding games often fails as projects grow because large context sizes degrade model performance. After evaluating Roblox MCP, Unity MCP, and open web libraries with Claude Code, the author found web tooling offered the strongest model proficiency but lacked high-level game abstractions. To bridge this gap, the author created VibeGame, an open-source declarative game engine built on three.js, rapier, and bitecs. VibeGame uses an Entity-Component-System architecture, XML-like markup, and bundled documentation in an llms.txt file to keep project context lean. While effective for basic physics and rendering mechanics, the engine requires further development for complex features like inventories, multiplayer, and interactions.
Swift Transformers Reaches 1.0 – and Looks to the Future
swift-transformers has released version 1.0 to provide a stable foundation for running local models on Apple Silicon platforms. The library supplies components missing from Core ML and MLX, focusing heavily on tokenization, chat templates, and Hugging Face Hub integration. In this major release, Tokenizers and Hub become standalone top-level modules, enabling developers to import only the components they require. The update also integrates an updated swift-jinja library for faster template processing, adopts Modern Core ML stateful models and MLTensor APIs, removes unnecessary CLI dependencies, and adds full Swift 6 support. Future development will prioritize deeper integration with MLX frameworks and agentic workflows like MCP.
Smol2Operator: Post-Training GUI Agents for Computer Use
Lightweight vision-language models often lack the precise visual grounding and action execution capabilities required for graphical user interface automation. To resolve this deficiency, the authors constructed a post-training pipeline that evolves base models into agentic GUI coders through supervised fine-tuning. The process standardizes disparate mobile and desktop dataset action spaces into a unified functional format utilizing normalized coordinates relative to image size. Training is structured into two sequential phases that first establish basic visual perception and subsequently introduce explicit reasoning steps with masked assistant loss. On the ScreenSpot-v2 perception benchmark, fine-tuning SmolVLM2-2.2B-Instruct increased element localization accuracy to 61.71 percent, while a compact 460-million-parameter model achieved approximately 58 percent accuracy.
SyGra: The One-Stop Framework for Building Data for LLMs and SLMs
Training and aligning large language models and small language models requires datasets tailored for complex reasoning, preference tuning, multi-turn questioning, and format conversion. To eliminate the need for bespoke data preparation scripts, SyGra provides a low-code and no-code Python framework for automated dataset creation, transformation, and alignment. The library integrates with diverse inference backends, including vLLM, Hugging Face TGI, Triton, and Ollama, allowing teams to focus on prompt engineering rather than pipeline infrastructure. It supports data workflows ranging from converting PDFs and knowledge bases into question-answering pairs to filtering low-quality samples and generating Direct Preference Optimization datasets. These plug-and-play workflows reduce manual curation effort while accelerating model fine-tuning and retrieval-augmented generation pipelines.
Gaia2 and ARE: Empowering the community to study agents
Existing AI agent evaluation environments are often tightly coupled to specific tasks and fail to model real-world challenges such as API failures, spontaneous events, and asynchronous conditions. To address this limitation, the Gaia2 benchmark and the Meta Agents Research Environments (ARE) framework introduce interactive read-and-write evaluation. Gaia2 incorporates 1,000 human-created scenarios spanning multi-step execution, cross-source search, ambiguity handling, adaptability, temporal reasoning, agent collaboration, and noise tolerance. Using a simulated smartphone interface equipped with 101 tools, evaluations showed that GPT-5 with high reasoning scored highest overall, while Kimi K2 was the leading open-source model. The results demonstrated that instruction following and search do not reliably predict performance on closer-to-real-world tasks.
Scaleway on Hugging Face Inference Providers 🔥
Scaleway is integrated as a supported serverless Inference Provider on the Hugging Face Hub, expanding model deployment options across Hub model pages and official client SDKs for JavaScript and Python. Operating out of European data centers located in Paris, France, Scaleway Generative APIs host open-weight models including gpt-oss, Qwen3, DeepSeek R1, and Gemma 3 with structured outputs, function calling, multimodal processing, and sub-200ms first-token response times. Developers can route inference requests using their Hugging Face tokens or supply direct Scaleway API keys. Billing for routed requests charges standard provider rates starting at €0.20 per million tokens without markups, while direct requests bill to Scaleway accounts. Hugging Face PRO subscribers receive two dollars of monthly inference credits applicable across providers.
Democratizing AI Safety with RiskRubric.ai
Cloud Security Alliance and Noma Security introduced RiskRubric.ai to provide standardized, transparent risk assessments across the open AI model ecosystem. The framework evaluates AI models across six pillars—transparency, reliability, security, privacy, safety, and reputation—using over 1,000 reliability tests, 200 adversarial security probes, automated code scanning, and harmful content evaluations. Each model receives 0–100 scores and A–F letter grades, supplemented by specific vulnerability findings and recommended mitigation strategies to assist deployment filtering. Initial benchmark results across models showed composite scores ranging from 47 to 94 with a median of 81, revealing polarized safety distributions and indicating that security hardening directly correlates with reduced safety risks.
Public AI on Hugging Face Inference Providers 🔥
Hugging Face has integrated Public AI as a supported Inference Provider on the Hugging Face Hub. Public AI operates as a nonprofit, open-source project providing access to sovereign and public models from institutions such as the Swiss AI Initiative and AI Singapore. Its distributed infrastructure combines a vLLM-powered backend serving OpenAI-compatible APIs across partner-donated clusters with a global load-balancing routing layer. Users can access these models through the Hugging Face web UI, Python client SDK, and JavaScript SDK using either direct provider API keys or routed Hugging Face tokens. At the time of announcement, inference through the Public AI provider is free of charge, supported by donated GPU time and advertising subsidies.
`LeRobotDataset:v3.0`: Bringing large-scale datasets to `lerobot`
LeRobotDataset:v3.0 is a new release for lerobot, Hugging Face’s robotics library, addressing file-system limitations encountered when LeRobotDataset:v2 stored one episode per file at million-episode scale. Version 3 packs multiple episodes into shared Parquet and MP4 files, while relational JSON and chunked episode metadata identify episode boundaries and locate the corresponding tabular and visual data. The format separates tabular data, visual data, and metadata; it stores low-dimensional readings in Apache Parquet, concatenates camera frames into MP4 videos, and supports native temporal windowing through delta_timestamps. It also introduces StreamingLeRobotDataset for processing batches from the Hugging Face Hub without downloading the full dataset locally, and provides a one-line conversion utility for datasets in the LeRobotDataset format. The release is planned for lerobot-v0.4.0 and is presented as a step toward datasets with millions of episodes.
Visible Watermarking with Gradio
Generative artificial intelligence tools produce images, video, audio, and text that have become increasingly difficult to distinguish from real-world media captures. To improve synthetic media transparency, Hugging Face introduced visible watermarking capabilities directly into the Gradio web application framework. Developers can now overlay visual watermarks on image and video outputs by specifying a single watermark parameter using file paths, open image objects, or NumPy arrays. The implementation also supports QR code watermarks, which can match the visual style of generated media while conveying detailed background information. Furthermore, the Gradio Chatbot component accepts a text watermark parameter that automatically attaches attribution metadata whenever users copy generated responses to their clipboards.
Introducing the Palmyra-mini family: Powerful, lightweight, and ready to reason!
WRITER released three open models in the Palmyra-mini family ranging between 1.5B and 1.7B parameters, designed for efficient inference and specialized reasoning tasks. Built on the Qwen architecture, the release includes the standard palmyra-mini base model alongside two Chain of Thought variants, palmyra-mini-thinking-a and palmyra-mini-thinking-b. In benchmark evaluations, palmyra-mini achieved 52.6% on Big Bench Hard, while palmyra-mini-thinking-a reached 82.87% on GSM8K and palmyra-mini-thinking-b reached 92.5% on AMC23. For palmyra-mini-thinking-b, applying reinforcement learning fine-tuning to an OpenReasoning-Nemotron-1.5B base improved single-shot pass@1 accuracy while reducing sampling diversity and majority@64 performance. The models are available in GGUF and MLX-BF16 quantizations and support inference engines including vLLM, SGLang, TRTLLM, and TGI.
Tricks from OpenAI gpt-oss YOU 🫵 can use with transformers
To support OpenAI's GPT-OSS series of models, the transformers library introduced several performance upgrades that apply across supported architectures. The release integrates zero-build custom kernels downloadable directly from the Hub, reducing external dependency bloat and compilation friction for operations like Liger RMSNorm and MegaBlocks MoE. Native support for MXFP4 quantization groups vector elements into 32-value blocks with shared scales, allowing GPT-OSS 20B to fit in roughly 16 GB of VRAM and GPT-OSS 120B in roughly 80 GB. In addition, transformers incorporates Flash Attention 3 with attention sinks, continuous batching via the generate_batch API for experimentation, and automatic memory pre-allocation to speed up model loading on GPUs.
Fine-tune Any LLM from the Hugging Face Hub with Together AI
Together AI and Hugging Face have introduced an integration enabling developers to fine-tune compatible Hugging Face Hub models directly on Together AI's managed infrastructure. To launch a fine-tuning job via the Python SDK, users provide a base model from Together's catalog as a configuration template alongside the target Hugging Face repository identifier. This base model template dictates GPU allocation, memory configuration, training pipelines, and inference setup for custom models with matching architectures and sizes. The workflow operates bidirectionally, pulling from public or token-authenticated private repositories and optionally pushing completed checkpoints back to the Hub upon completion. Geared toward CausalLM models under 100 billion parameters, the capability enables faster iteration cycles and domain adaptation without custom DevOps infrastructure.
Jupyter Agents: training LLMs to reason with notebooks
Small language models often struggle to compete with large frontier models on complex, agentic data science tasks. To improve notebook-based reasoning, researchers simplified agent scaffolding down to roughly two hundred lines of code with dedicated execution and final answer tools, boosting baseline easy accuracy on the DABStep benchmark from 44.4 percent to 59.7 percent. They constructed a curated training dataset by deduplicating two terabytes of Kaggle notebooks, automatically fetching five terabytes of linked datasets, and scoring educational value and relevance using Qwen3-32B. Question-answer pairs grounded in verified execution traces were generated to fine-tune compact Qwen3-4B thinking and instruct models. The team released the trained models alongside the dataset and execution sandboxes, establishing a foundation for reinforcement learning and distillation on notebook workflows.
mmBERT: ModernBERT goes Multilingual
mmBERT is a massively multilingual encoder model trained on more than 3T tokens across over 1,800 languages to improve upon existing multilingual architectures like XLM-R. Built upon ModernBERT with a Gemma 2 tokenizer, mmBERT employs a three-phase training curriculum consisting of pre-training on 60 languages, mid-training on 110 languages, and a final decay phase covering 1,833 languages. The training pipeline integrates an inverse mask ratio schedule, dynamic language temperature annealing, and TIES merging across three decay variants. Benchmark evaluations demonstrate strong natural language understanding on English GLUE and multilingual XTREME, as well as competitive retrieval performance on MTEB v2 and CoIR. The release includes standard base and small models alongside open training data and checkpoints.
Welcome EmbeddingGemma, Google's new efficient embedding model
Google DeepMind released EmbeddingGemma, a multilingual embedding model with 308 million parameters and a 2048-token context window designed for on-device applications. Based on the Gemma3 transformer backbone, the architecture replaces causal attention with bidirectional attention to function as an encoder, followed by mean pooling and two dense layers producing 768-dimensional vectors. The model incorporates Matryoshka Representation Learning, allowing outputs to be truncated down to 512, 256, or 128 dimensions for reduced memory and storage footprints. Trained on approximately 320 billion multilingual tokens across more than 100 languages, the quantized model operates under 200 MB of RAM. In domain-specific evaluations on the MIRIAD dataset, fine-tuning increased NDCG@10 from 0.8340 to 0.8862, outperforming larger baselines.
SAIR: Accelerating Pharma R&D with AI-Powered Structural Intelligence
SandboxAQ released the Structurally Augmented IC50 Repository (SAIR), an open-source dataset containing 5.24 million computationally co-folded 3D protein-ligand structures paired with empirical IC50 binding potency data. The repository addresses training data scarcity in structure-based drug discovery, where over 40 percent of the included target proteins lack experimental structures in the Protein Data Bank. To construct the dataset, engineers executed over 130,000 GPU hours of the Boltz1 co-folding model on 760 NVIDIA H100 processors hosted via NVIDIA DGX Cloud on Google Cloud Platform. Infrastructure optimizations maintained over 95 percent GPU compute utilization, compressing the generation timeline from three months to three weeks. Quality validation using PoseBusters confirmed that 97 percent of the predicted complexes met physical plausibility and chemical sanity standards.
Policy Gradient with PyTorch
Policy gradient algorithms optimize reinforcement learning policies directly without learning intermediate action-value functions. Instead of assigning discrete Q-values to actions, these methods parameterize a stochastic policy and apply gradient ascent on an objective score function. This setup eliminates manual exploration tuning, resolves perceptual aliasing in identical states, and naturally accommodates continuous or high-dimensional action spaces. The Reinforce Monte Carlo policy gradient algorithm operationalizes this approach by collecting full episodic trajectories and adjusting parameters along the gradient of the log-action probabilities scaled by return. Practitioners implement this algorithm using PyTorch to evaluate agent robustness across environments like CartPole-v1, PixelCopter, and Pong.
Liftoff! How to get started with your first ML project 🚀
Beginners in machine learning frequently face difficulties when selecting a framework and defining the scope of their initial hands-on project. Sentence Transformers provides an accessible starting point by computing dense vector representations for sentences, paragraphs, and images, enabling semantic search through cosine similarity. A structured four-step strategy guides practitioners through listing library capabilities, identifying compelling datasets, selecting a familiar secondary tool, and brainstorming project concepts. Applying this framework, a song lyrics dataset was paired with Gradio Blocks to build a prompt-based playlist generator. Developing this application required evaluating pre-trained models, hosting generated embeddings on Hugging Face Spaces, and utilizing multi-processor support to accelerate embedding generation.
Accelerate Large Model Training using DeepSpeed
Training large models on hardware with limited GPU memory frequently causes out-of-memory errors when using standard Distributed Data Parallel. To solve this bottleneck, Hugging Face Accelerate integrates DeepSpeed ZeRO data parallelism to shard optimizer states, gradients, and model parameters across workers. For a 900-million-parameter DeBERTa model, ZeRO Stage 2 increased the maximum per-device batch size from eight to forty while achieving a 3.5-fold training speedup over DDP without degrading accuracy or F1 score. Advanced configurations allow sequence-to-sequence chatbot finetuning and full ZeRO Stage 3 CPU offloading for models like the 1.5-billion-parameter GPT-XL, enabling batch size 16 training where DDP fails. Accelerate enables these memory optimizations through simple configuration files and the accelerate launch command with minimal or zero code modifications.
Topics they write about
All topics →Weekly digest
The best engineering reads, once a week.