Appearance

The Trending Open Source ML Tools Actually Used In Production Right Now

June 20, 2025 6 min read
#open-source #ml-infrastructure #llm-training #developer-tools #privacy

If you check GitHub trending right now, you will not see a new 1T parameter model. You will not see a new benchmark leader.

You will see boring tools. Tools that solve specific, annoying, unglamorous problems that every ML team hits when they stop demoing and start running things in production.

This is not an accident. We have passed the peak of model novelty. For most teams the hard problem is no longer what model to use. It is how you train it, clean the data going into it, redact PII coming out of it, get training data in the first place, and not go bankrupt on cloud bills doing any of this.

This article breaks down every top trending ML and developer tool from the last 30 days, what they actually do, the parts they leave out of the readme, and who should be using them.

Over the last 30 days these are the most starred, forked and actively contributed open source repositories used by ML engineers. None are base models. All solve production problems.

This is not a curated list. This is exactly what is being adopted right now, in order of actual activity:

  1. Unsloth
  2. slime
  3. Presidio
  4. yt-dlp
  5. public-apis
  6. Home Assistant
  7. geo-seo-claude

Every one of these has earned their position. None are viral hype. All have clear, measurable advantages over every alternative.

Unsloth: Everyone is fine tuning on this now ​

Unsloth is a LLM training and inference runtime. It is currently the default choice for every independent developer, startup and most mid sized teams doing fine tuning.

The numbers are not marketing. For standard LoRA fine tuning you will get ~2x speedup and ~70% lower VRAM usage with zero accuracy loss. For GRPO reinforcement learning that gap opens to 2x speed and 80% less VRAM.

This is not achieved with clever marketing. It is achieved with hand written Triton kernels for every hot path in the transformer, and zero abstraction overhead. Unsloth does not wrap Hugging Face Transformers. It replaces the parts that are slow.

You can run it via a one line install script, or via their new Studio UI which works natively on Windows, Linux and macOS. It supports every major model released in the last 18 months. The team works directly with model authors to fix bugs upstream before public release.

There are caveats. Multi node training is still immature. The Studio UI is AGPL licensed, the core library remains Apache 2.0. Support for AMD and Intel GPUs exists but is 1-2 months behind NVIDIA.

None of that matters. If you are fine tuning any LLM today and you are not using Unsloth, you are paying twice as much for half the speed. That is the entire argument.

slime: The RL framework that all the SOTA models actually use ​

slime is an LLM post training framework for reinforcement learning scaling. It is the single most important new ML infrastructure release of 2025, and almost no one outside the frontier model teams has heard of it.

Every major open RL model release from the last 6 months was trained on slime. This includes GLM 5, DeepSeek R1, Qwen 3.5, and every derivative that beat MMLU or HumanEval.

slime does one thing very well. It connects Megatron-LM for training directly to SGLang for rollout, with a shared data buffer and explicit data flow. There is no abstraction layer. Every argument for Megatron and SGLang is passed through directly.

This is the opposite design choice of every previous RL framework. All prior frameworks tried to support every training backend and every inference engine, and ended up being slow, buggy and missing all the good features of each backend. slime picks exactly one good training stack and exactly one good rollout stack, and optimizes the connection between them.

It has full end to end tests. It has explicit tracing for silent RL bugs. It supports delta weight sync, disaggregated prefill/decode, and heterogeneous GPU clusters.

If you are trying to run RL training at any scale larger than a single 4090, stop using OpenRLHF. Stop using TRL. Use slime.

Presidio: The only PII redactor you will ever need ​

Presidio is a context aware PII de-identification service for text and images. It is maintained by Microsoft, fully open source, and used internally at almost every large company running LLMs.

It supports every common PII type out of the box: credit cards, names, locations, social security numbers, phone numbers, bitcoin addresses, financial identifiers. It combines regex, checksum validation, context aware NER and rule based logic. You can add custom recognizers in 10 lines of code.

You can run it as a Python library, a Spark UDF, a Docker container or a Kubernetes service. It will correctly redact PII that generic NER models miss, and will almost never produce false positives on technical text.

There is no good alternative. Every other open source PII redactor is either unmaintained, produces garbage output, or is a wrapper around Presidio anyway.

If you are accepting user input into an LLM, or outputting LLM generated text to a user, you should be running Presidio. There is no excuse not to.

yt-dlp: The most important open source project no one talks about ​

yt-dlp is a command line audio and video downloader. It supports over 3000 sites. It is also the single most depended upon open source repository in the entire ML ecosystem.

Virtually every multi modal training dataset built in the last three years uses yt-dlp. Every fine tuning dataset for video, every speech recognition dataset, every demonstration dataset for agent models pulls data using this tool.

It is maintained almost entirely by volunteers. It survives constant breakage from sites changing their APIs, DRM and anti bot measures. It has perfectly backwards compatible command line interface dating back to the original youtube-dl release in 2006.

No one funds it. No one gives it awards. Almost every major AI company in the world uses it every single day.

If you are building any kind of multi modal system, you will use yt-dlp. You should donate to the maintainers.

public-apis: Stop building scrapers ​

public-apis is a manually curated list of 1800+ public APIs sorted by category, with standardised metadata for authentication, HTTPS support and CORS status.

This repository is 10 years old. It is still updated every single day. It is the single fastest way to check if something you are about to scrape already has a public API.

Every entry includes authentication requirements, HTTPS status, CORS support and working links. Categories cover every domain from weather and finance to government open data, machine learning and test data.

Before you write a single line of scraper code, check this list. 9 times out of 10 someone has already done the work.

Home Assistant: The privacy blueprint ​

Home Assistant is open source home automation software. It is included here not because ML engineers are all building smart homes, but because it is the single best working example of privacy first open source infrastructure that exists today.

It puts local control first. It never phones home. It has zero mandatory cloud services. It supports 3000+ devices. It runs on a Raspberry Pi. It is maintained by a community, not a venture backed company.

This is the blueprint for how all ML infrastructure should be built. Right now almost every LLM tool does the exact opposite. They phone home by default. They require cloud accounts. They lock you into their platform.

If you are building developer tools, go look at Home Assistant. Copy what they do.

geo-seo-claude: The first tool built for the AI search collapse ​

geo-seo-claude is an audit tool for Generative Engine Optimization. It is the first practical tool built explicitly for the world where half of all search traffic never reaches your website.

AI search already drives 527% year over year traffic growth. It converts at 4.4x the rate of traditional organic search. Only 23% of marketers are actively optimizing for it.

This tool runs 12 separate audits against a domain, including AI crawler access, citation readiness, brand signal scanning and llms.txt validation. It produces actionable scores and client ready reports. It runs entirely as a Claude Code skill with no hosted backend.

If you run any kind of public website, you will be running these audits before the end of this year. This tool is currently the only way to do it properly.

The quiet trend no one is commenting on ​

Look again at this entire list.

None of these projects were founded by venture backed startups. None have raised 100M dollar rounds. None have marketing teams. None announce releases on Twitter.

All of them won by just being better. All of them solve specific, boring problems. All of them got popular by word of mouth between engineers.

This is the shift happening right now in open source ML. After 3 years of hype cycles, demo videos and benchmark gaming, engineers are finally defaulting back to tools that just work.

Tools that do one thing. Tools that don't break. Tools that don't try to lock you in. Tools that are honest about their limitations.

What this means for your workflow ​

You can ignore 99% of announcements on Hacker News. You can ignore every new base model release. You can ignore every benchmark leaderboard.

If you want to actually build things, use these tools.

Fine tune with Unsloth. Run RL with slime. Redact PII with Presidio. Get data with yt-dlp and public-apis. Optimize for AI search with geo-seo-claude. Copy the privacy model from Home Assistant.

That is the entire production ML stack for 2025. Everything else is noise.

References ​

All source repositories referenced:

Support Me