90% of ML repositories that hit GitHub trending are abandoned 3 days later. Most are wrapper scripts, demo gifs with no working code, or forks with one line changed and a viral tweet attached.

This week was different. Six projects broke the top 10. Four of them will change how you build systems this quarter. Two you can ignore entirely.

The bar for this list

Every repository included here hit trending in the last 72 hours. None require an API key to test. All have permissive open licenses. None made the list just for having a pretty demo video.

This is not a curation of cool things. This is a list of things you will probably end up running in production this year.

TimesFM: The first usable general purpose forecasting foundation model

Google Research dropped this without fanfare on Tuesday. It is the most important release this month.

TimesFM is a 200M parameter time series foundation model pretrained on 100B real world time points across retail, finance, server metrics, IoT and industrial sensor data. It runs zero shot forecasting for 1 to 512 step horizons. No fine tuning is required for 90% of common business use cases.

Benchmark results are not subtle. On the M5 competition test set, TimesFM achieves 12.7% lower MAE than Prophet, 8.2% lower than NeuralProphet, and comes within 1.1% of the winning hand tuned competition entries. That is an absurd result for a model that has never seen the target dataset.

It has limitations. It chokes on very strong seasonal signals with periods longer than 1 year. It does not handle missing values well. License is Apache 2.0. It runs unquantized on a 16GB T4.

If you are building any forecasting pipeline this quarter, stop what you are doing and test this first.

Anthropic Cybersecurity Skills: The first standard agent capability library

Almost nobody is talking about this repository correctly. This is not a prompt list. This is not a demo.

This is 754 structured, machine readable cybersecurity skills mapped 1:1 to every major industry security framework: MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND and NIST AI RMF. Every skill has defined inputs, expected outputs, success criteria, failure modes and validation checks.

This is the first common capability interface that every major agent runtime already supports. It works out of the box with Claude Code, GitHub Copilot, Cursor, Gemini CLI and 17 other platforms. License is Apache 2.0.

Right now every team building security agents is reinventing this exact list from scratch. This repository just removed 6 months of work for every single one of them. This will be a dependency in 90% of enterprise agent projects by the end of the year.

LongLive 2.0: Nvidia fixed long video infrastructure

Everyone posted the 10 minute demo clips. Almost nobody looked at the actual code release.

LongLive 2.0 is not just another video model. It is the first training and inference pipeline that can generate consistent video over 10 minutes at 30fps without identity drift, scene collapse or repeated motion artifacts. The architecture uses per scene latent caching and incremental rolling attention windowing.

Nvidia released full training code, not just inference weights. There are no hidden dependencies. There are no usage restrictions beyond the standard Nvidia source license.

This is the new baseline. Ignore every other open video generation repository for the next month. Everyone will be copying this implementation.

Databricks AI Dev Kit: The boring agent toolkit that works

Databricks field engineering dropped this unannounced. There is no blog post. There is no press release.

This is the exact toolkit Databricks engineers use when they build agents for enterprise customers. It has production reference implementations for tool calling retries, context window eviction, hallucination checking, audit logging and error handling. All the boring, unglamorous parts that no demo repository ever includes.

There is no lock in. This runs on any cloud, any base model. License is Apache 2.0.

Virtually every public agent repository shows you how to build a hello world chatbot. This one shows you how to build one that does not break when you give it to real users.

The ones you can skip

Not everything trending is worth your time.

free-claude-code is a proxy wrapper that routes requests through a public third party endpoint. It will get rate limited or shut down within 10 days. Do not build anything on top of this. Do not enter any credentials.

ai-engineering-from-scratch is at time of writing nothing but a very good README and a promise. There is zero code. Star it. Come back in a month. Do not waste an evening on it right now.

What this trend tells us

Notice what is missing. There are no new base LLMs. There are no new transformer architectures. There are no papers claiming a new state of the art on MMLU.

All of the useful work this week is infrastructure, standards and tooling.

That is where the field is right now. Everyone has stopped arguing about which base model is best. Everyone has stopped posting one line benchmarks. Everyone is now trying to actually build things that work reliably.

This is not a temporary lull. This is the end of the base model gold rush. For the next 12 months, almost all meaningful progress will happen in repositories that look exactly like the ones on this list.