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
Security incident disclosure — July 2026
An intrusion into part of production infrastructure was driven end to end by an autonomous AI agent system, accessing a limited set of internal datasets and several credentials used by services. The attack began through two code-execution paths in dataset processing, then escalated from a worker to node-level access and moved laterally across internal clusters while running thousands of actions in short-lived sandboxes. AI-assisted anomaly detection surfaced the compromise, and LLM-driven analysis of more than 17,000 attacker events reconstructed the timeline, identified indicators, mapped credentials, and separated impact from decoys. Hosted frontier models initially blocked forensic requests containing attack artifacts, so the investigation used zai-org/GLM-5.2 on internal infrastructure. The initial vulnerabilities were closed, affected infrastructure rebuilt, credentials rotated, and cluster controls and alerting strengthened; assessment of possible partner or customer impact remains ongoing.
Model Routing Is Simple. Until It Isn’t.
Model routing in agentic systems is presented as a systems optimization problem rather than a simple classification task. On 417 AppWorld Test Challenge tasks with the same CodeAct agent, Claude Sonnet 4.6 cost $79 total ($0.19 per task), versus $155 ($0.37 per task) for GPT-4.1, despite GPT-4.1 having lower listed token prices; cache-read pricing and repeated context changed the effective result. The discussion also shows why task difficulty and model speed are incomplete signals: hidden execution work, specialization, governance constraints, routing overhead, cache state, hardware, and endpoint load all affect decisions. The described router optimizes cost, quality, and latency together, producing configurable cost-accuracy operating points; one latency-oriented configuration reached 84% accuracy for $93 and 83 seconds, while adding about 6 ms and 2 kB per task.
Welcome Inkling by Thinking Machines
Thinking Machines Lab has released Inkling and the smaller Inkling-Small family as open multimodal models that accept image, text, and audio inputs. Inkling is described as a roughly 1T-parameter decoder-only Mixture-of-Experts model, with 975B total and 41B active parameters, a 1M-token context window, and training on 45 trillion multimodal tokens. Its architecture combines relative attention, alternating global and sliding-window attention, short convolution, shared experts, hierarchical image patchification, and discretized mel-spectrogram audio processing. The release includes BF16 and NVFP4 checkpoints, speculative MTP layers, day-zero support in Transformers, SGLang, vLLM, and llama.cpp, plus deployment configurations and interactive voice and image demos. Reported benchmarks show Inkling and Inkling-Small achieving strong but varied results across reasoning, coding, vision, audio, factuality, chat, and safety evaluations.
Introducing Real World VoiceEQ: Measuring the human quality of voice AI
Existing voice-AI benchmarks increasingly suggest near-human performance, while real-world conversations still expose failures involving speaker consistency, hesitation, accents, noise, emotion, and natural responses. Real World VoiceEQ measures human quality in voice interaction by testing whether systems recognize, produce, and respond to acoustic information omitted by transcripts, including tone, emotion, speaker identity, and background context. It evaluates more than 40 proprietary and open-source models across 15+ dimensions and 60+ metrics covering ASR, TTS, S2S, and Speech Understanding, using more than 1 million human ratings, including 785,000 TTS and 48,000 STS ratings. Findings indicate that capabilities are specialized, speech-to-speech systems often underuse paralinguistic cues, and traditional benchmarks can overestimate performance in noise and other conversational conditions. The research also reports that human agreement remains essential for subjective acoustic and social judgments because SLM evaluators were less reliable outside clear, verifiable tasks.
Shipping huggingface_hub every week with AI, open tools, and a human in the loop
huggingface_hub, the Python client underlying libraries such as transformers, datasets, diffusers, and sentence-transformers, moved from releases every four to six weeks to a weekly cadence. A single manually triggered GitHub Actions workflow now computes versions, manages branches, tags and publishes both huggingface_hub and the hf CLI, creates downstream RC test branches, and handles post-release tasks. An open-weights model drafts structured release notes and Slack announcements using PR metadata and documentation diffs, while OpenCode, HF Inference Providers, PyPI Trusted Publishing, and pinned, checksum-verified tooling provide the stack. Deterministic manifest validation checks missing or extra PR references, then a human reviews and publishes the drafts. The reported result is weekly shipping at about $0.25 per full release, with more consistent notes, earlier downstream breakage detection, and shorter contributor feedback loops.
Experimenting with the proposed Cross-Origin Storage API in Transformers.js
Transformers.js browser apps currently redownload and recache identical AI model and WebAssembly resources when they are requested from different origins. Chrome's cache uses a Network Isolation Key composed of the top-level site and current-frame site, so matching CDN URLs do not produce cache hits; the toy example incurs 177 MB of duplicate model storage, while a shared ONNX Runtime Wasm file is 4,733 kB. The proposed Cross-Origin Storage (COS) API addresses this by identifying large files with cryptographic hashes and sharing them across opted-in origins, with a polyfill extension enabling experimentation before native browser support. In the extended Transformers.js flow, a COS lookup can serve resources in milliseconds, while a miss downloads and stores them, and ModelRegistry.is_pipeline_cached() can help select among models already cached.
PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M Parameters
PP-OCRv6 is PaddleOCR’s latest universal OCR model family, targeting text detection and recognition across documents, screenshots, displays, labels, and scene text. It offers tiny, small, and medium tiers spanning 1.5M to 34.5M parameters, while the small and medium tiers support 50 languages. Its PPLCNetV4 backbone is shared across detection and recognition, with RepLKFPN for lightweight multi-scale detection and EncoderWithLightSVTR for recognition using local context and global attention. On PaddleOCR’s official in-house multi-scenario benchmarks, PP-OCRv6_medium reaches 86.2% detection Hmean and 83.2% recognition accuracy, improving over PP-OCRv5_server by 4.6 percentage points and 5.1 percentage points. PaddleOCR provides Paddle Inference, Transformers, and ONNX Runtime paths, and results can be saved as visualization images or structured JSON for downstream workflows.
We got local models to triage the OpenClaw repo for FREE!*
OpenClaw receives hundreds of issues and pull requests daily, creating a triage problem for maintainers who need near-real-time filtering without consuming a paid closed-model quota. The system uses Gemma and Qwen local models inside a Pi-based agent harness: each item is normalized, enriched with GitHub context, classified into finite topic labels, and required to return structured output. For pull requests, the agent can inspect the repository through reposhell, a read-only command-restricted shell, while localpager, SQLite, a worker queue, and deterministic notification policy handle orchestration and Discord delivery. Tests described for gemma-4-26b-a4b and qwen3.6-35b-a3b report hundreds of tokens per second and good zero-shot classification accuracy without fine-tuning, while an LLM audit loop monitors errors; the authors present the pattern as agentic classification for high-throughput triage.
MosaicLeaks: Can your research agent keep a secret?
MosaicLeaks examines privacy leakage in deep-research agents that combine private enterprise documents with external web retrieval. Queries that appear harmless individually can reveal sensitive facts when an observer combines them over multiple hops, producing intent, answer, or full-information leakage. The benchmark contains 1,001 multi-hop chains interleaving local and web questions, and compares baseline behavior with task-only reinforcement learning and Privacy-Aware Deep Research (PA-DR). Task-only training raises strict chain success from 48.7% to 59.3% but increases answer/full-information leakage from 34.0% to 51.7%, whereas PA-DR reaches 58.7% success with 9.9% leakage by rewarding stage-specific decisions and penalizing query-level privacy risk. The evaluation is controlled, using synthetic enterprise documents, a fixed web corpus, three company contexts, and one multi-hop question-answering harness.
Beyond LoRA: Can you beat the most popular fine-tuning technique?
Parameter-efficient fine-tuning (PEFT) reduces the memory required to adapt open models, enabling fine-tuning with a fraction of the memory and, in some cases, quantized models. LoRA is dominant because it freezes base weights and trains a small set of added parameters, but its popularity may also reflect ecosystem momentum and uneven comparisons. The post presents shared-condition benchmarks for language-model math reasoning and image generation, using identical models, datasets, code, hardware, and measures including test performance, VRAM, forgetting or drift, runtime, and checkpoint size. The results show that LoRA performs well but is not consistently best: other techniques can outperform it on one or more axes, especially in the image-generation benchmark. The recommended choice is therefore to compare techniques against a use case’s needs and constraints, including quantization support, layer coverage, adapter merging, and downstream serving compatibility.
Is it agentic enough? Benchmarking open models on your own tooling
Final-answer checks can show whether an agent completed a task, but they miss the turns, tokens, seconds, errors, and command path required. The proposed tool-specific benchmark measures that work on transformers, comparing models, revisions, tasks, and three access tiers: bare installation, source clone, and a packaged Skill containing CLI documentation and examples. It uses deterministic tasks with exact matching, launches one Hugging Face Job per model–revision–task combination on identical hardware, and records traces in a Hugging Face Bucket. Reports include match rate, median new, cached, and generated tokens, time, error rate, and marker adoption, while traces expose the agent's commands. On transformers, the CLI plus Skill helped large open models but caused Qwen3-14B to misinterpret documentation as an executable tool, dropping classify-sentiment match from 100% to 0% in the cited comparison.
From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot
Strands Robots integrates Hugging Face’s LeRobot stack with Strands Agents to connect dataset recording, policy execution, simulation, hardware deployment, and fleet coordination in one agent loop. Its integration keeps LeRobot responsible for hardware drivers, calibration, recording, and dataset handling, while Strands supplies AgentTools, a shared Robot abstraction, policy providers, and a Zenoh peer mesh; simulation and hardware use the same LeRobotDataset format. The walkthrough records demonstrations in MuJoCo, optionally pushes them to the Hub, runs policies through a common interface, and switches from simulation to a physical SO-101 with mode="real". The companion script and notebook run the default simulation workflow on a laptop without hardware, a GPU, or Hugging Face credentials, although the default Mock policy produces structurally valid but behaviorally unhelpful random demonstrations.
GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2 is presented as a flagship model for long-horizon tasks, extending GLM-5.1’s capability to a solid 1M-token context for sustained engineering work. It was trained extensively on 1M-context coding-agent scenarios involving large-scale implementation, automated research, performance optimization, and complex debugging, and adds effort-level controls to balance capability, latency, and computational cost. Its architecture uses IndexShare, sharing one lightweight indexer across every four transformer layers and reducing per-token FLOPs by 2.9× at 1M context; MTP changes for speculative decoding increase acceptance length by up to 20%. On reported long-horizon and standard coding benchmarks, GLM-5.2 is described as the highest-ranked open-source model: it trails Opus 4.8 by 1% on FrontierSWE, ranks second only to Opus 4.8 on PostTrainBench, and scores 81.0 on Terminal-Bench 2.1 versus GLM-5.1’s 63.5.
Agentic Resource Discovery: Let agents search
Agentic Resource Discovery (ARD) is a draft, open specification for discovering agent capabilities at runtime, addressing a model in which users and developers must pre-install, integrate, and maintain each MCP tool, skill, or A2A agent. It defines a static ai-catalog.json manifest and a dynamic POST /search registry API, using indexed signals such as publisher identity, representative queries, compliance attestations, and tags to move capability selection outside the LLM. Hugging Face's Discover Tool adapts Hub semantic search into ARD catalog entries for running Spaces, exposing skills, MCP servers, or raw Space metadata according to the requested media type and wrapping agents.md files into installable SKILL.md artifacts. The tool is available through the hf CLI, REST, and MCP, while planned work includes federation modes and Hub support for profile-level manifests.
Profiling in PyTorch (Part 2): From nn.Linear to a Fused MLP
Part 2 of the series examines how PyTorch profiler traces change from a single bias-enabled nn.Linear to a three-layer MLP with an intervening activation, using scripts run on an NVIDIA A100-SXM4-80GB GPU. It explains that nn.Linear represents y = x @ w.T + b, dispatches aten::addmm, and folds bias addition into the cuBLAS GEMM epilogue, while aten::t only rewrites weight strides and launches no GPU kernel. For the MLP, eager execution uses three GEMMs plus separate activation and multiplication kernels, whereas torch.compile fuses the latter operations and keeps the intermediate out of HBM, while leaving the GEMMs unchanged. A hand-written Liger Triton kernel provides the same fusion without Dynamo guards or compile latency, trading Inductor’s shape-specialized speed for generic execution across changing shapes.
How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces
A coding agent built a 3D Paris gallery by chaining two Hugging Face Spaces: an image-generation Space produced dark-background monument images, and VAST-AI/TripoSplat reconstructed each single image as a 3D Gaussian splat. It then handled integration work by correcting the splats’ Y-down orientation, auto-framing monuments, compressing .ply files to .ksplat at about one-third the size, and building and deploying a static Three.js viewer. The Spaces’ agents.md files provided schemas, call and polling templates, file-upload instructions, and authentication guidance, allowing the agent to call them without a client library. The same pipeline produced six-monument galleries for Egypt and Japan after the prompts changed. The account presents callable, documented multimedia components as a way to reduce integration work and make new multimedia applications cheaper to create.
Migrating Your GitHub CI to Hugging Face Jobs
Trackio's GitHub Actions CI was limited by generic hosted machines, maintenance-related delays, and lack of readily available GPU runners, while the project needed CPU checks and CUDA-based tests. The setup keeps GitHub Actions as the CI coordinator and uses a dispatcher Space plus the huggingface/jobs-actions bridge to launch ephemeral GitHub runners inside Hugging Face Jobs, selected through labels such as hf-jobs-cpu-upgrade and hf-jobs-t4-small. The guide covers duplicating and configuring the dispatcher, creating a GitHub App, supplying runner credentials and an HF token, changing runs-on, choosing Docker images, and validating runs and logs; Playwright and NVIDIA CUDA images are used for CPU and GPU examples. In Trackio, the CPU job took 1m10s versus a 1m40s GitHub ubuntu-latest baseline, while the GPU check passed in 45s and cost less than a cent at the t4-small rate; logs were available through both systems.
The Open Source Community is backing OpenEnv for Agentic RL
OpenEnv is becoming a community-coordinated, open-source tool for building agentic execution environments such as terminals and browsers, with the stated goal of helping train open-source agents. Starting now, a committee including Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, Microsoft, Hugging Face, and RadixArk will coordinate the project, which now lives at huggingface/OpenEnv and is supported or adopted by numerous AI organizations. Technically, OpenEnv is being narrowed to an interoperability and protocol layer: environments expose a Gymnasium-style reset(), step(), and state() API through client/server deployments, standard HTTP or WebSocket protocols, Docker packaging, and MCP compatibility. It does not define rewards or training loops; those remain in specialized libraries, while OpenEnv connects harnesses, environments, and trainers across ecosystems. Planned work includes external rewards, dataset-backed tasksets, harness integrations, end-to-end examples, and automatic environment validation.
Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI
Nemotron 3.5 Content Safety is NVIDIA’s 4B-parameter model for evaluating user prompts, optional images, and optional assistant responses together, with multilingual coverage and enterprise-specific policy enforcement. It extends Nemotron 3 by accepting natural-language custom policies, producing optional auditable reasoning traces, and supporting binary verdicts, category-labelled verdicts, or THINK mode; disabling THINK preserves the default low-latency path. Built on Google Gemma 3 4B IT with a LoRA adapter and a 128K context window, it uses the Aegis 2.0 taxonomy of 13 core categories and 10 subcategories, and targets deployment on GPUs with 8GB or more of VRAM. The accompanying multimodal, multilingual dataset includes concise reasoning traces, while reported results average 96.5% on Multilingual Aegis, 88.8% on RTP-LX, and about 85% across the evaluated benchmark set.
Designing the hf CLI as an agent-optimized way to work with the Hub
The post presents a redesign of hf, the Hugging Face Hub’s official CLI, for both human users and coding agents that download, upload, manage repositories, run Jobs, and administer other Hub resources. The CLI detects agents through environment variables, switches to compact, complete TSV output, and supports explicit human, agent, JSON, and quiet formats; agent mode was introduced in hf v1.9.0. Next-command hints, stderr separation, non-interactive failure messages, retry-safe operations, confirmations, and dry-run previews support multi-step automation without removing human-friendly terminal behavior. Benchmarks report equal-or-better success with roughly 1.3× to 1.8× fewer tokens than curl or the SDK on comparable tasks, and up to 6× fewer on complex workflows; the generated hf-cli skill reduced tool calls by about 30% without reducing token use or improving reliability.
Topics they write about
All topics →Weekly digest
The best engineering reads, once a week.