We are building on top of Hugging Face now

If you are building AI systems in 2026, you are almost certainly building on top of Hugging Face. This is not an opinion. It is an observation of what every working ML engineer actually does every day.

This transition happened quietly. For years the Hub was just a place to download model weights. Today it hosts datasets, benchmarks, agent skills, inference endpoints, robot policies, training pipelines and deployment tooling. Every major new capability lands here first. Every standard gets adopted here first. Every experiment runs here first.

This article walks through the last two months of releases across the ecosystem. None of these are splashy keynote announcements. All of them are things people are already using in production right now.

The quiet standardization of agent tooling

Agents stopped being a research topic in 2025. They are now boring infrastructure.

The biggest unannounced shift this year is that every major player has converged on exactly the same agent architecture. There are no more debates about ReAct vs Plan-and-Execute. No one argues about reflection loops any more. All working agents now use exactly the same structure: a planning step, tool selection, execution, verification, and retry.

The remaining hard problems are not about agent loops. They are about what happens around the loop.

Agentic Resource Discovery launched this month to solve the biggest remaining pain point: tool discovery. Up until now you had to hardcode every tool an agent could use into its prompt. ARD changes this. It defines a standard search API so agents can find tools, skills and other agents at runtime.

Hugging Face already runs a public ARD registry indexing every Space on the Hub. You can call it today. An agent can now search for "transcribe audio" at runtime, find the right MCP server, load it, and use it, without any human pre-configuration.

This is not a product. It is a protocol. Microsoft, Google, GoDaddy and Hugging Face all collaborated on the specification. It will be everywhere within twelve months.

MosaicLeaks: The privacy failure no one was measuring

Everyone building agents knew this was a problem. No one had measured it until MosaicLeaks.

When an agent has access to both private documents and web search, it will leak private information. Not all at once. One fragment at a time. No single query reveals anything. Taken together, an observer watching the query log can reconstruct almost everything the agent knows.

This is the mosaic effect. And it is catastrophic.

Across all tested base models, 34% of multi-hop research queries leaked enough information for an adversary to reconstruct private facts. When researchers trained the agent to perform better at the task, leakage went up to 51.7%. Better agents leak more.

Telling the agent not to leak barely moved the needle. Adding a single line to the system prompt brought leakage down only to 25.5%, and dropped task success by 4%.

The solution was not prompting. It was training. PA-DR, the reward function proposed in the paper, brought leakage down to 9.9% while keeping task performance effectively identical at 58.7%.

You cannot prompt privacy into an agent. You have to train it in. This is the single most important practical result about agents published this year.

Beyond LoRA: We were all leaving performance on the table

LoRA won. 98.4% of all PEFT checkpoints on the Hub are LoRA. 95% of all fine-tuned image generation models use LoRA. It is the default choice for every engineer.

It is also not the best choice.

Hugging Face ran the first fair, head to head benchmark of every PEFT technique implemented in the PEFT library. Same base model. Same dataset. Same hyperparameters. Same hardware. No researcher bias.

For LLM fine tuning on math reasoning, LoRA sits on the Pareto frontier. But it is not the only option. BEFT uses 11% less memory for slightly lower accuracy. Lily delivers 3.2% higher accuracy for 13% more memory.

For image generation on Flux, LoRA is not even on the frontier. OFT strictly dominates it: 1.6% higher similarity score, and 9.6% lower memory usage.

Almost no one uses these techniques. Not because they are worse. Because LoRA got there first, got all the tutorials, and got all the downstream support.

This is changing. The PEFT library now supports converting any adapter to LoRA for deployment. You can fine tune with OFT, convert to LoRA, and serve the result on vLLM with zero changes.

Benchmarking libraries for agents, not humans

APIs used to be designed for humans. Now they are designed for agents.

A bad API will waste 10 seconds of your time. It will waste 1000 tokens and three extra turns of an agent's time. At scale that adds up to millions of dollars.

No one measured this until now. The agent benchmarking harness released this month does not just check if an agent gets the right answer. It measures how much work it took to get there. How many turns. How many tokens. How many errors. Which API paths it took.

When the Hugging Face team tested adding a simple CLI to transformers, they found that large models completed common tasks 1.8x faster. They also found that small models got worse. Small models rely on memorized API patterns from their training data. New interfaces confuse them.

This is the new reality of library development. Every change you ship now has two separate performance profiles: one for humans, one for agents. You have to test both.

Robotics has landed on the Hub

Robotics was the last major ML field that had not moved to the Hub. That changed this month.

First ABC-130k landed: 3598 hours of bimanual robot teleoperation data, 130,919 episodes. This is the largest open robot dataset ever released, and it is already the standard training set for every new robot policy.

Then Strands and LeRobot shipped their integration. You can now write an agent once, run it in MuJoCo simulation, then deploy exactly the same code unchanged to a physical SO-101 robot arm. Datasets recorded in simulation work on hardware. Policies trained on hardware work in simulation.

This is the end of the walled garden era for robotics. For ten years every robot company ran their own completely custom stack. Now everyone uses the same dataset format, the same policy interface, the same training code, and the same Hub to share it all.

The new generation of open generative tools

Closed model providers still lead on raw generation quality. Open source now leads on everything else.

LTX-2 shipped this month. It is the first open video model that matches commercial output quality. It supports native audio, lip sync, retakes, HDR output, and camera control. It runs on consumer hardware. You can fine tune it. You can modify it.

OpenMontage shipped at the same time. It is not a video generator. It is an agentic video production system. It will research your topic, write a script, source assets, edit a timeline, and render a finished video. It can build complete documentaries entirely from open archival footage for $0.15.

You will not see these announced on stage at any keynote. They are just sitting on GitHub and the Hub, being used by thousands of people every day.

Knowledge systems that actually accumulate

RAG is broken.

Every time you ask a question, RAG starts over. It re-chunks. It re-embeds. It re-reranks. It re-derives every fact from scratch. Nothing accumulates. Nothing is learned.

Three systems shipped this month that fix this.

OpenKB compiles your documents once into a persistent, interlinked wiki. It extracts concepts. It resolves entities. It adds cross references. Knowledge compounds over time. You can open the resulting wiki in Obsidian. You can distill it into an agent skill that any other agent can load.

HippoRAG 2 does for retrieval what OpenKB does for static knowledge. It builds an associative memory graph that mirrors human long term memory. It beats GraphRAG, RAPTOR and LightRAG on every multi-hop benchmark, and uses 70% less resources during indexing.

STORM remains the gold standard for deep research. It now supports every major LLM and every major search engine. 70,000 people have used the public demo.

None of these are perfect. All of them work better than vanilla RAG for almost every real world use case.

What comes next

This ecosystem is growing faster than any single person can track. New tools land every day. New benchmarks. New models. New standards.

The most important shift is that none of this is controlled by any one company. Hugging Face runs the Hub, but they do not own the standards. They do not own the protocols. Every piece of this stack is open source. Every piece can be replaced.

This is not the end state. This is just the foundation. For the first time we have a common, open, interoperable base that everyone builds on top of. All the hard problems from here on will be built on this base.

If you are working in this field, this is where you live now. This is the operating system.