This is not another trending Github list.

Every single tool covered here is being used daily by senior ML engineers right now. None are demo repos. None have fake launch marketing. Every one solves a specific, universal pain point that every ML team hits, and every one works well enough that people will stop what they are doing to install it in the middle of a work day.

We are at an interesting inflection point. For the first time, most of the best tools for ML engineering are open source, not commercial. The commercial vendors are now playing catch up.

The state of play

There are 9 tools here, across 7 completely separate domains. None compete with each other. You can install all of them on the same machine today and they will all make your work easier.

Kronos: Stop training your own financial time series model

Everyone working in quant finance has wasted 6 months building exactly this model. Kronos is the first foundation model built explicitly for candlestick K-line data, trained across 45 global exchanges. It is not a general time series model. It does not pretend to work for weather or server metrics. It does exactly one thing, and it does it better than anything else available.

Model	Context length	Params	Open source
Kronos-mini	2048	4.1M	✅
Kronos-small	512	24.7M	✅
Kronos-base	512	102.3M	✅
Kronos-large	512	499.2M	❌

You can get a working forecast for any asset in 7 lines of code. There is a proper fine tuning pipeline, proper batching, proper normalization handling. It was accepted to AAAI 2026, and every quant shop I know is running fine tunes on this right now.

This is the model that killed 90% of the internal time series foundation model projects at hedge funds this year.

Outlines: Structured outputs done correctly

Every single person building with LLMs has this problem. You ask for JSON. You get JSON with a trailing comma. You add 3 paragraphs of instructions. You get markdown wrapped JSON. You write a regex parser. The LLM changes the output format the next day.

Outlines solves this completely. It guarantees valid structure during generation, not after. It works with every model, every provider, and uses native Python types. No custom DSL. No magic prompts.

review = model(product_review_text, ProductReview)

That is the entire API. You pass a Pydantic model, you get back a valid instance of that model. 100% of the time. It works for enums, unions, literals, functions. There is no catch.

This is the single most installed LLM utility library of 2026. If you are still parsing LLM outputs with regex, stop what you are doing and install this now.

SkyPilot: The GPU orchestrator everyone switched to

SkyPilot is the tool that won the workload orchestration war. Everyone was using Slurm. Everyone was complaining about Kubernetes for ML. Everyone tried 12 different commercial GPU schedulers. Now almost everyone is using SkyPilot.

It does one very simple thing extremely well: you write one task definition once, and you can run it on any GPU, any cloud, any Kubernetes cluster, any Slurm cluster. It will automatically find the cheapest available instance, fail over, sync code, setup dependencies, stream logs.

Shopify runs all their AI training on it. Multiverse doubled their GPU utilization after switching. It has native agent integration so your coding assistant can spin up an 8x A100 cluster for you while you are writing code.

There is no longer any good reason to use anything else for running ML workloads.

Agent Reach: Give your agent internet access that actually works

Agents are great at code. They are absolutely terrible at the internet. Ask them to read a twitter thread, they will tell you the API is paid. Ask them to summarize a youtube video, they will give you the title. Ask them to check reddit for a bug, they get 403 blocked.

Agent Reach fixes this. It is not another scraping API. It is a capability layer that maintains the current working method to access every major platform. When yt-dlp gets blocked by Bilibili, they switch the backend. When twitter kills the anonymous API, they switch to OpenCLI. You don't change anything. You just tell your agent to install it, and 2 minutes later it can read every site on the internet.

This is the single most important utility for anyone using coding agents right now.

Browser Use: Agents that can actually use a browser

Browser Use is the current best in class browser automation agent. It scores 87.4% on the Odysseys benchmark, ahead of every commercial agent from OpenAI, Anthropic, Google and Microsoft.

You can install it into Claude Code, Cursor, or any other agent with one line. After that you can say things like:

Upload this video to youtube Fill out this job application with my resume Compare these three laptops and make me a price table

And it will just do it. It clicks buttons, types into forms, handles captchas, navigates pagination. It works exactly like a human would use a browser.

This is not a research demo. People are using this every day to do actual work.

TileLang: Write GPU kernels in Python that beat hand written assembly

Writing high performance GPU kernels was a black art reserved for 100 people on the planet. That changed this year.

TileLang is a DSL built on top of TVM that lets you write GPU kernels with normal Python syntax, that perform within 2% of hand optimized assembly. It supports NVIDIA, AMD, Apple Metal, Ascend NPUs. You can write a full Flash Attention implementation in 80 lines of code.

This is the biggest improvement to kernel developer productivity in the last 10 years. Every major inference engine is porting their kernels to TileLang right now.

Memanto: Persistent memory that works for agents

Every agent has amnesia. You spend 3 hours explaining your codebase to it. The context resets. You have to start over.

Memanto fixes this. It is an active memory agent, not a vector database. It remembers everything you told the agent, across sessions, across restarts, across different tools. It runs entirely locally. No API keys required.

It scores 89.8% on LongMemEval, ahead of Mem0, Zep and Letta. You install it, run one command to connect it to your agent, and you will never have to explain the same thing twice.

PyTorch Geometric: The GNN library that won

There used to be 12 different graph neural network libraries. Now there is one. PyTorch Geometric.

It has every single published GNN layer implemented correctly. It has proper sampling for large graphs. It has proper multi GPU support. It compiles with torch.compile. It is maintained, it is stable, it works.

If you are doing any work with graphs, this is the only library you need.

AI Engineering From Scratch: The curriculum everyone is using

84% of students use AI tools. Only 18% feel prepared to use them professionally. This curriculum closes that gap.

It is 20 phases, 503 lessons, 320 hours. It starts with linear algebra. It ends with multi agent swarms. You build every single component from scratch before you are allowed to use the framework version. By the time you get to PyTorch you already know exactly what it is doing under the hood.

This is now the standard onboarding curriculum for ML teams at half the major AI companies.

Closing observations

None of these tools came out of big companies. None of them had 10 million dollar launch marketing budgets. All of them were built by small teams of engineers solving problems they themselves had.

This is the new normal for ML tooling. The best tools are no longer built by vendors. They are built by the people using them.

If you are not using these tools yet, you are working harder than you need to.