✷ The archive · 11 labs · 871 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
Profiling in PyTorch (Part 3): Attention is all you profile
Hugging Face’s Profiling in PyTorch (Part 3) shows how different attention implementations appear in PyTorch profiler traces, revealing performance trade‑offs of naive, in‑place, math, efficient, flash, and cuDNN backends.
Hugging Face Native-speed vLLM Transformers Modeling Backend
Hugging Face has updated the transformers vLLM backend to match or exceed the throughput of custom vLLM implementations by dynamically applying inference-specific layer fusions at runtime.
Hugging Face and Amazon SageMaker Studio Integration
Hugging Face has introduced a deep-link integration with Amazon SageMaker AI, allowing developers to move from model discovery to fine-tuning or deployment in SageMaker Studio with a single click.
Hugging Face Models on Foundry Managed Compute
Microsoft Foundry now integrates a curated, weekly-refreshed catalog of Hugging Face open-weight models that can be deployed in one click onto Foundry Managed Compute for enterprise-grade operationalization.
Hugging Face and SkyPilot Integration for Zero-Egress AI Storage
Hugging Face and SkyPilot have integrated to allow AI workloads to run on any cloud provider with zero-egress costs for reading models and datasets stored on the Hugging Face Hub.
LeRobot v0.6.0 release notes / what's new
LeRobot v0.6.0 introduces world model policies, a expanded VLA model zoo, a unified reward models API, and a new deployment CLI with DAgger-style human-in-the-loop corrections.
PRX Data Strategy: Scaling Pre-training with VLM Re-captioning and Mosaic Streaming
Photoroom details the data pipeline for PRX, emphasizing the use of long VLM-generated captions, a hybrid Lance and Mosaic Data Shards storage strategy, and high-quality JPEG encoding to optimize a 7B parameter model.
Hugging Face Kernels Major Updates
Hugging Face announced major updates to its Kernels project, introducing a new kernel repository type, trusted publishers and code signing, revamped CLIs, broader framework support, and foundations for agentic kernel development.
Hugging Face and Cerebras Real-Time Voice AI with Gemma 4
Hugging Face and Cerebras have developed an open, cascaded speech-to-speech pipeline using Gemma 4 31B to enable natural, low-latency voice AI interactions.
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
IBM Research introduces ScarfBench, an open benchmark to evaluate AI agents' ability to migrate enterprise Java applications across Spring, Jakarta EE, and Quarkus frameworks.
Why Specialization Is Inevitable in AI Systems
Based on the 2026 research by Goldfeder et al., specialization is a structural necessity for AI performance because finite resources make concentrated capacity more effective than broad generality.
Hugging Face Community Evals and Every Eval Ever (EEE) Integration
Hugging Face has integrated Community Evals with the Every Eval Ever (EEE) project to enable cross-posting of standardized evaluation results and direct linking between model pages and detailed technical records.
DiScoFormer: One transformer for density and score, across distributions
DiScoFormer is a new transformer-based model that estimates both the density and score of a distribution from a set of data points in a single forward pass without requiring retraining for new distributions.
Hugging Face Jobs: Deploying vLLM Servers with a Single Command
Hugging Face now allows users to deploy private, OpenAI-compatible vLLM endpoints on its infrastructure using a single command via HF Jobs, providing a pay-per-second alternative for testing, evaluations, and batch generation.
NVIDIA NeMo AutoModel Accelerates Fine-Tuning of Mixture-of-Experts Models
NVIDIA NeMo AutoModel provides 3.4-3.7x higher training throughput and 29-32% lower GPU memory usage for fine-tuning Mixture-of-Experts models by building on Hugging Face Transformers v5 with Expert Parallelism, DeepEP, and TransformerEngine kernels, requiring only a one-line import change.
Hugging Face FFASR Leaderboard: Benchmarking Far-Field ASR
Hugging Face and Treble Technologies have launched the FFASR Leaderboard, the first open community-driven benchmark to quantify the performance gap between near-field and far-field Automatic Speech Recognition (ASR) in realistic acoustic environments.
Experimenting with the Cross-Origin Storage API in Transformers.js
Hugging Face shows how Transformers.js can use the experimental Cross-Origin Storage API to cache model and Wasm resources by hash, eliminating duplicate downloads across origins.
Hugging Face huggingface_hub Release Automation
Hugging Face has transitioned from a 4-6 week release cycle to a weekly cadence for huggingface_hub by implementing an AI-driven, human-in-the-loop CI/CD pipeline using open-source tools and open-weights models.
PP-OCRv6 release: 50-Language OCR from 1.5M to 34.5M Parameters
PaddlePaddle has released PP-OCRv6, a scalable OCR model family supporting 50 languages with parameter counts ranging from 1.5M to 34.5M.
Hugging Face Local Models for OpenClaw PR Triage
Hugging Face demonstrates how local models like Gemma 4 and Qwen 3.6, deployed in an agentic harness, can perform real-time, cost-free triage of GitHub issues and pull requests for the OpenClaw repository.
MosaicLeaks: Addressing Privacy Leakage in Deep Research Agents
Hugging Face introduces MosaicLeaks, a benchmark and the Privacy-Aware Deep Research (PA-DR) training method to prevent research agents from leaking private enterprise data through their external web queries.
Hugging Face Benchmarking Open Models on Agentic Tooling
Hugging Face introduces a new benchmarking harness to evaluate how different model sizes and library revisions affect the efficiency and success rate of coding agents using software tools.
Hugging Face PEFT: Evaluating Alternatives to LoRA
Hugging Face's benchmarking of the PEFT library reveals that while LoRA is widely popular, other parameter-efficient fine-tuning techniques like OFT and Lily can outperform it in memory efficiency and test accuracy depending on the task.
Strands Robots and LeRobot Integration: From Hugging Face Hub Datasets to Physical Robot Deployment
Hugging Face announced the Strands Robots SDK integration with LeRobot, enabling users to record robot demonstrations, push them to the Hub, run policies in simulation, and deploy the same code to physical SO-101 robots with a single argument change, while coordinating multiple robots via a Zenoh-based mesh.
GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2, released by Z.AI on Hugging Face, introduces a solid 1M-token context, improved architecture via IndexShare and MTP enhancements, effort-level control, and strong open-source performance on long-horizon coding benchmarks.
Agentic Resource Discovery (ARD) Specification and Hugging Face Implementation
Hugging Face has launched a reference implementation of the Agentic Resource Discovery (ARD) specification, an open standard that allows AI agents to dynamically search for and integrate tools, skills, and other agents at runtime.
Profiling in PyTorch (Part 2): From nn.Linear to a Fused MLP
Hugging Face explains how to optimize Multilayer Perceptrons (MLPs) in PyTorch by analyzing GPU kernel fusion, the impact of torch.compile, and the use of hand-tuned kernels via the kernels library.
Hugging Face Spaces agents.md enables automatic 3D multimedia pipelines
A coding agent automatically built a static Hugging Face Space that showcases Paris monuments as 3‑D Gaussian splats by chaining two existing Spaces, demonstrating how `agents.md` turns any Space into a composable building block for multimedia pipelines.
Migrating GitHub CI to Hugging Face Jobs
Hugging Face introduces a method to run GitHub Actions on serverless Hugging Face Jobs infrastructure, enabling GPU-accelerated CI and reducing CPU job runtimes by up to 30%.
OpenEnv: Standardizing Agentic RL Environments for Open Source AI
Hugging Face has transitioned OpenEnv to a community-coordinated project to serve as an interoperability layer for agentic reinforcement learning environments.
NVIDIA Nemotron 3.5 Content Safety Release
NVIDIA released Nemotron 3.5 Content Safety, a 4B-parameter multimodal, multilingual safety model with custom policy enforcement and optional reasoning traces for enterprise AI.
Hugging Face hf CLI agent‑optimized redesign cuts token usage and boosts success rates
Hugging Face announced an agent‑optimized redesign of the hf CLI that cuts token usage by up to 6× and improves success rates for coding agents like Claude Code and Codex.
Direct Preference Optimization Beyond Chatbots
The provided source material is a 429 rate limit error page from Hugging Face and contains no technical content regarding Direct Preference Optimization.
Hugging Face Rate Limit Prevents Access to 'Adding MCP Tools to Reachy Mini' Post
The Hugging Face link to the blog post 'Adding MCP Tools to Reachy Mini' returns a 429 rate limit error, so no content is available to summarize.
Holo3.1: Fast & Local Computer Use Agents
The provided source material for Holo3.1 is unavailable due to a 429 rate limit error, and no technical details were provided.
JetBrains Mellum2 Release
JetBrains has announced the release of Mellum2, a 12B Mixture-of-Experts (MoE) model, though the source material provided is an error page and contains no technical details.
Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic – Summary
Hugging Face’s post on “Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic” could not be retrieved, so no technical details can be summarized.
Profiling in PyTorch: A Beginner's Guide to torch.profiler
Hugging Face provides a comprehensive guide to using torch.profiler to identify bottlenecks, understand the CPU-GPU dispatch chain, and analyze the impact of torch.compile on kernel execution.
Reachy Mini Local Speech Backend Integration
Hugging Face has released a local speech-to-speech pipeline for Reachy Mini, allowing the robot to handle conversations fully locally using a cascaded VAD, STT, LLM, and TTS stack.
Delta Weight Sync in TRL Enables Trillion-Parameter Model Training with Minimal Bandwidth
Hugging Face announced Delta Weight Sync in TRL, a feature that reduces weight synchronization bandwidth in async RL training by over 100x by transmitting only sparse weight changes via Hugging Face Buckets, enabling disaggregated training without shared clusters.
Hugging Face AI Agent Glossary: Defining Harness, Scaffold, and Agent Architecture
Hugging Face provides a standardized vocabulary for AI agents, defining the agent as the combination of a model, a harness for execution, and scaffolding for behavior definition.
OlmoEarth v1.1 release notes / what's new
Hugging Face and AllenAI have released OlmoEarth v1.1, a family of Earth observation models that reduces compute costs by up to 3x while maintaining performance similar to v1.
Ettin Reranker Family v1 release notes
Hugging Face released the Ettin Reranker family, six CrossEncoder rerankers from 17M to 1B parameters distilled from mxbai-rerank-large-v2, achieving state-of-the-art retrieval reranking quality and speed.
PaddleOCR 3.5 release notes / what's new
PaddleOCR 3.5 introduces Hugging Face Transformers as a supported inference backend, allowing OCR and document parsing models to integrate more seamlessly into PyTorch-based workflows.
Granite Embedding Multilingual R2 Release: Open Apache 2.0 Multilingual Embeddings with 32K Context
IBM Granite released two Apache 2.0 multilingual embedding models—granite-embedding-97m-multilingual-r2 (97M) and granite-embedding-311m-multilingual-r2 (311M)—with 32K-token context, 200+ language support, and top sub-100M retrieval scores.
Unlocking asynchronicity in continuous batching
Hugging Face shows how to overlap CPU and GPU work in continuous batching using CUDA streams and events, cutting LLM inference time by about 22% for an 8B model generating 8K tokens.
Building Blocks for Foundation Model Training and Inference on AWS
Hugging Face details a four-layer architectural framework on AWS—comprising infrastructure, resource orchestration, ML software stacks, and observability—to support the evolving scaling laws of pre-training, post-training, and test-time compute.
vLLM V1 Migration: Ensuring Backend Correctness in Reinforcement Learning
ServiceNow AI achieved parity between vLLM V0 and V1 for RL rollout generation by fixing logprob semantics, runtime defaults, weight update paths, and implementing an fp32 lm_head.
Hugging Face Open ASR Leaderboard: Private Datasets to Combat Benchmaxxing
Hugging Face has introduced private evaluation datasets from Appen Inc. and DataoceanAI to the Open ASR Leaderboard to prevent test-set contamination and provide a more robust measure of real-world ASR performance.
IBM Granite 4.1 LLMs release notes / technical overview
IBM has released Granite 4.1, a family of dense, decoder-only LLMs (3B, 8B, and 30B) that achieve high performance through rigorous data curation and a multi-stage reinforcement learning pipeline.