The AI Morning Post
Artificial Intelligence • Machine Learning • Future Tech
Lead Story
Transformers Crosses 164,000 Stars as Hugging Face's Library Becomes the De Facto OS of Modern ML
Hugging Face's Transformers library has quietly become the most starred production ML framework on GitHub, a milestone less about a single breakthrough and more about the consolidation of an entire ecosystem around one repository.
The transformers library, once a convenient wrapper for BERT-style models, has evolved into something closer to an operating system for machine learning: a common substrate that lets a researcher swap in a new architecture — audio, vision, or a DeepSeek-style reasoning model — without rewriting the surrounding scaffolding. Its climb past 164,000 stars, with 34,200 forks, reflects less a single viral moment and more years of steady accretion as labs from DeepSeek to smaller open-source shops standardize on its APIs.
This matters because infrastructure dominance often precedes model dominance. PyTorch, sitting at 102,400 stars just below it on today's trending list, tells the same story from a different layer of the stack: the tensor engine underneath, largely invisible to end users, yet indispensable to nearly every model release covered in this newsletter. Together the two repositories represent the plumbing through which almost all contemporary AI research now flows.
The quieter implication is consolidation risk. When one library and one runtime become this central, bugs, deprecations, or governance decisions inside Hugging Face and Meta's PyTorch team ripple outward to the entire field. The next architectural leap in AI may well be gated not by a lab's compute budget, but by how quickly a single open-source maintainers' team can review a pull request.
By the Numbers
Deep Dive
The Long Tail of Hugging Face: What Zero-Download Repos Tell Us About the Platform's Real Economy
Scroll through today's Hugging Face trending list and a strange pattern emerges: five repositories, ranked by momentum, and not one of them has meaningful download or like counts. A voiceprint model with zero engagement sits at rank one. A vision-tuned checkpoint, a metacognitive chain-of-thought experiment, and a sprint-themed personal project round out the list. This is not an anomaly — it is the platform working as designed, and it reveals something important about how attention on Hugging Face actually functions.
Unlike GitHub stars, which accumulate slowly and require a user to take a deliberate action, Hugging Face's trending algorithm rewards velocity — the rate of change in traffic, clones, and page views over a short window. A brand-new repository with a handful of curious visitors can outrank an established model with thousands of steady downloads, simply because its growth curve is steeper in relative terms. This produces a trending page that often looks less like 'what's important' and more like 'what's new and being poked at right now.'
The consequence is a bifurcated ecosystem. On one side sit the load-bearing repositories — the base models, the widely-forked frameworks — that quietly accumulate millions of downloads without ever needing to trend. On the other sit thousands of small, often single-author uploads: personal fine-tunes, research checkpoints, merged models built with tools like TIES or DARE, many of which will never be downloaded outside their creator's own experiments. Today's list, dominated by names like 'iamseungpil' and 'danielharkin21' rather than institutional labs, is a snapshot of this second economy — the individual tinkerer's layer of Hugging Face that rarely gets analyzed but makes up the platform's overwhelming numerical majority.
For journalists and analysts, the lesson is to treat 'trending' as a measure of curiosity, not consequence. The real signal of influence remains cumulative downloads, sustained citation, and adoption inside production pipelines — metrics that move slowly and reward the transformers and PyTorchs of the world, not the flashiest new upload of the week.
Opinion & Analysis
We Need Better Words Than 'Trending' for What's Actually Happening
Every platform that surfaces momentum — GitHub, Hugging Face, app stores — inherits the same problem: velocity metrics conflate genuine breakthroughs with statistical noise from tiny baselines. A repository going from zero to five views can 'trend' the same way a repository going from ten thousand to fifty thousand does.
Until platforms start reporting confidence intervals or absolute engagement thresholds alongside their trending badges, readers — and reporters — will keep mistaking a ripple for a wave. It's a small fix, but one that would meaningfully improve how the industry understands its own attention economy.
Model Merging Isn't Cheating, But It Is a Different Kind of Progress
Techniques like TIES-merging, seen today in a 35B-parameter Qwen variant, let anyone combine multiple fine-tuned models into a new checkpoint without additional training. Critics call this a shortcut; I'd call it a different research question entirely — not 'how do we train better models' but 'how do we combine the models we already have more intelligently.'
Both questions matter. But the ML community should be honest that a merged model's benchmark gains often say more about which base models were chosen than about any new capability being unlocked. Attribution matters, and right now the ecosystem's citation norms haven't caught up with its merging tools.
Tools of the Week
Every week we curate tools that deserve your attention.
Ultralytics YOLO26
Latest YOLO release unifying detection, segmentation, and classification in one lightweight package
TIES-Merge Toolkit
Popular method for combining fine-tuned LLM checkpoints without retraining from scratch
scikit-learn 1.6
The classic ML library keeps shipping — still the fastest path from CSV to baseline model
Keras 3.0
Multi-backend deep learning framework now running seamlessly across JAX, TensorFlow, and PyTorch
Trending: What's Gaining Momentum
Weekly snapshot of trends across key AI ecosystem platforms.
HuggingFace
Models & Datasets of the WeekGitHub
AI/ML Repositories of the Week🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text
Tensors and Dynamic neural networks in Python with strong GPU acceleration
A curated list of awesome Machine Learning frameworks, libraries and software.
scikit-learn: machine learning in Python
Deep Learning for humans
Ultralytics YOLO26, YOLO11, YOLOv8 — object detection, instance segmentation, semantic segmentation,
Biggest Movers This Week
Weekend Reading
The Bitter Lesson, Ten Years Later
Rich Sutton's essay on compute-over-cleverness remains the best lens for reading today's infrastructure consolidation
Understanding Model Merging: TIES, DARE, and Beyond
A technical primer on why merged checkpoints are proliferating across Hugging Face
The Economics of Open Source Infrastructure
Essential context for why a handful of repos like transformers and PyTorch end up governing an entire industry
Subscribe to AI Morning Post
Get daily AI insights, trending tools, and expert analysis delivered to your inbox every morning. Stay ahead of the curve.
Join Telegram ChannelScan to join on mobile