This is not a roundup of announcement press releases. This is what working engineers actually uploaded to public Hugging Face repositories between May 14 and May 21 2026.

None of these artifacts had launch blog posts. None were tweeted by CEOs. None have made it onto any AI news site. Every single one is immediately usable for production work.

The artifact breakdown

This week there were 12 high signal artifacts that received real usage. Not demo toys, not benchmark fine tunes, not throwaway forks.

Artifact	Type	Author	Total interactions	Notes
Ornith 1.0 35B GGUF	Model	deepreinforce-ai	11247	First general purpose 35B that runs acceptably on 24GB VRAM
Unlimited OCR	Model + Space	Baidu	7892	No character limit, no rate limits, works on rotated handwritten text
ITBench-AA	Dataset	ArtificialAnalysis	3106	SRE incident diagnosis benchmark
HIW-500	Dataset	BitRobot	48869	500 hours of humanoid robot teleoperation
Wan 2.2 FP8 Preview	Space	r3gm	1890	First public working demo of Wan 2.2 video generation
Pro Realism Edit Studio	Space	Sneak-Moose	219	Photorealism correction for generated images
LOCUS-v1	Dataset	LocalLaws	912	Structured US municipal law corpus
open-perfectblend	Dataset	mlabonne	2460	Curated fine tuning dataset
arxiv-latex	Dataset	scholarweave	6426	Full parsed arXiv LaTeX corpus
Krea 2 Turbo	Model	krea	1741	Fast image generation
Boogu-Image	Space	multimodalart	65	Open source Midjourney clone
Cuga Apps	Space	ibm-research	44	GPU utilization optimizer demo

The end of the standalone demo

Nobody builds standalone demo sites any more.

Every single one of the new models released this week had an official Hugging Face Space up before any announcement was made. Most had working demos 6+ hours before anyone posted about the model anywhere else.

Zero setup. Zero auth. Zero rate limits for reasonable usage. You click once and it runs on Hugging Face hardware. This is now the expected standard. If you release a model and do not post a working Space within 2 hours, 90% of people will ignore your release entirely.

This flow is now completely standard. It takes less than 10 minutes end to end. Nobody waits for peer review. Nobody waits for press. Nobody waits for anything.

Ornith 1.0: the local 35B that actually works

deepreinforce-ai dropped two Ornith variants this week: 9B and 35B, both quantized to GGUF.

This is the first 35B parameter model that will run at acceptable speed on a consumer RTX 4090. At Q4_K_M quantisation it uses 21.7GB VRAM and generates 38 tokens per second. It beats Llama 3.1 70B on 7 out of 12 common benchmarks. It beats GPT-4o Mini on every coding benchmark.

Nobody announced this. There was no thread on X. It just appeared. It has already been downloaded 11,000 times.

Unlimited OCR: Baidu just reset the bar

Baidu uploaded Unlimited OCR on Tuesday.

There is no fine print. No page limit. No character limit. Works on scanned documents, handwritten notes, screenshots, rotated text, partially obscured text. It outputs structured markdown, preserves layout, extracts tables correctly. It beats every commercial OCR API including Google Cloud Vision and AWS Textract on every independent test.

It is fully open weight. You can run it locally. You can embed it in production software. There are no usage restrictions.

This is the single most useful open model released in the last 6 months. Almost nobody has noticed yet.

Datasets that people use, not just cite

For years 99% of datasets uploaded to Hugging Face existed only to be cited in papers. That changed this month. All of the top downloaded datasets this week are built for production training, not benchmark scores.

ITBench-AA

This is a dataset of 102 real Kubernetes production incidents. Each entry includes full cluster state, alerts, traces and the verified root cause. Agents are given read only access and must correctly identify the fault.

Every SRE team on earth is currently running their internal agent against this dataset. It is the only benchmark that measures something that actually matters. Generic LLM leaderboards do not matter any more. Performance on ITBench is the only number that engineering leads care about right now.

HIW-500

BitRobot dropped 500 hours of human teleoperation data for the Unitree G1 humanoid. 23,000 episodes recorded in 12 real homes. Full joint state, stereo camera feed, action traces, subtask annotations.

Before this dataset, all public humanoid training data was recorded in clean lab environments. This is the first real world dataset. Every humanoid robotics team on earth downloaded this within 12 hours of upload. That is why it has 48,000 downloads.

arxiv-latex

scholarweave fixed the single most annoying problem in academic ML: getting the actual LaTeX source for arXiv papers.

The official arXiv bulk bucket is requester pays. A full download costs $450 in egress fees. This dataset is parsed, cleaned, structured and hosted for free. You can download the entire thing for $0. It already has 6,400 downloads.

The broken dataset viewer bug nobody talks about

You may have noticed the HIW-500 dataset viewer returns a cast error. This is not an accident.

The Hugging Face dataset viewer does not correctly handle nested structs with more than 7 levels of nesting. Any dataset with properly structured sensor telemetry will break the viewer completely.

This bug has existed for 11 months. It has 17 open issues on GitHub. It will not be fixed any time soon. For production datasets, the viewer being broken is now a signal that the dataset contains real, high resolution structured data, not toy CSV rows.

What this tells us

Hugging Face stopped being a research hub a long time ago. It is now the default global distribution layer for production ML artifacts.

Nobody announces things any more. They just upload them. If you are waiting for press releases or blog posts you are 2-4 weeks behind everyone else.

The most important work happening right now is not being announced. It is just being uploaded.

Production usage notes

• All Ornith quantizations work correctly with llama.cpp, ollama and text-generation-webui. The Q5_K_M variant is the best balance of speed and quality.
• Unlimited OCR will use ~7GB VRAM at 16 bit. There is an unofficial 4 bit quantize that runs on 4GB.
• HIW-500 is 10TB total. You can download individual episodes instead of the full dataset.
• ITBench evaluation scripts are in the dataset files. Most models score between 12% and 38% correct root cause identification.
• The Wan 2.2 demo will generate 7 second 1080p video in approximately 90 seconds on Zero hardware.