This is not a list of demo repos.

Every tool covered here hit GitHub trending this month, has active maintainers, production deployments, and solves a problem that teams were paying enterprise license fees for 6 months ago. None require you to sign up for a waitlist. None ask for your email before you can run code.

The quiet end of demo culture

For almost three years every trending ML repo was a wrapper that made a pretty screenshot and broke on the third request. That era is over. Every repo on this list ships working code, has benchmark numbers, and includes the ugly edge case handling that separates toys from production tools.

No one is building general purpose AGI demos anymore. People are building tools that solve specific problems. That is good news for everyone who actually runs code in production.

AI Berkshire: The first useful agent framework

Everyone has built an agent demo. No one has built an agent that people actually trust with high stakes decisions. Until now.

AI Berkshire is not a general purpose agent. It is a set of structured Claude Code skills that implement value investment methodology with enough rigor that the author has run real money through it for two years.

Benchmark	2024 Full Year	2025 YTD
AI Berkshire live portfolio	+69.29%	+66.38%
S&P 500	+23.31%	+16.39%
Hang Seng	+17.67%	+27.77%
NASDAQ	+28.64%	+20.36%
CSI 300	+14.68%	+17.66%

This is the first open source agent that passes the most important test: it is used to make decisions that cost real money.

The architecture is not clever. That is the point.

Most agent frameworks try to make agents agree. This one is designed so that they disagree. That is the entire secret. Buffett will say something is cheap. Li Lu will say no one knows what it will look like in 10 years. That tension is exactly what good decision making looks like.

MinerU: Document parsing that beats commercial tools

Document parsing is the most boring, most expensive, most broken part of every RAG pipeline. Everyone hates it. No one talks about it.

MinerU is an open source document parser that now outperforms every commercial parser except Adobe Extract on most benchmarks. And it runs fully offline.

Version 3.4 released this month doubled OCR speed, added native PPTX and XLSX parsing, and dropped AGPL dependencies. It runs on CPU, every domestic AI chip, and will process 1000 page documents without running out of memory.

If you are currently paying for any commercial document parsing API you should test this this week. For 95% of use cases it will be better.

Free LLM API Resources: The operational reference every engineer needs

This repo is just a list. It is also the most useful thing published on GitHub this quarter.

It tracks every legitimate free LLM API endpoint, their actual rate limits, token quotas, and fine print. No reverse engineered endpoints. No scams. No hidden terms that disable the free tier after 3 days.

Provider	Free Monthly Token Quota	Top Free Model
Groq	1,000,000,000	Llama 3.3 70B
Google Gemini	36,000,000	Gemini 3.5 Flash
OpenRouter	30,000,000	GPT OSS 120B
Cloudflare Workers AI	10,000,000	Qwen 3.6 27B
Cohere	7,200,000	Command A+

Do not run benchmarks against OpenAI anymore. Every model on this list is available for free right now. You can run 100k test runs for zero cost.

The maintainer updates this list daily. Bookmark it.

NeMo: NVIDIA finally shipped good open source speech

NeMo pivoted hard this quarter. It is no longer a general purpose model toolkit. It is now the best open source speech stack available, by a very wide margin.

The new Nemotron 3.5 ASR runs 2400 concurrent streams on one H100 at 80ms latency. It supports 40 languages. It beats OpenAI Whisper large v3 on WER across every tested latency profile.

This is not a research model. This is production ready code that you can deploy today. There is no reason for anyone to run Whisper for new deployments anymore.

LlamaIndex: It got boring, that is good

LlamaIndex stopped shipping flashy demo features six months ago. Instead they fixed bugs, standardized interfaces, and stabilized the core. That is exactly what you want from a framework you deploy to production.

The import namespace cleanup is done. The plugin system works. Most importantly, they stopped breaking API changes every minor release.

If you are building any kind of RAG or agent system today, LlamaIndex is the boring, reliable default choice. LangChain is still the more popular choice for demos. LlamaIndex is what people use when they need something that runs every night without someone on call.

PyTorch: The baseline everyone takes for granted

PyTorch hits trending every single month. No one comments on it. No one writes blog posts about it. It just works.

This month they merged full CUDA 13.2 support, fixed the last major torch.compile bugs, and made source builds work reliably with uv.

There is no competition. There will not be competition. Every other framework discussed in this article runs on top of PyTorch. It is the foundation.

AutoGPT grew up

AutoGPT invented the agent hype cycle. Then it disappeared for 18 months while everyone else copied the demo.

They came back this month with an actual production platform. No more infinite loop demos. It has proper workflow orchestration, lifecycle management, monitoring, and a standard agent protocol that actually works with other tools.

It is still rough around the edges. But it is no longer a toy.

What changed

All of these tools have one thing in common: none of them try to be general purpose.

Every one solves one specific, narrow, extremely annoying problem very well. None claim to be AGI. None have animated hero sections. None have a Discord server full of people asking when it will be able to write code.

This is what maturity looks like. The hype cycle is over. We are now building tools.

Closing observation

Twelve months ago every one of these capabilities would have been a $50/month SaaS product. Today they are open source, maintained, and production ready.

The cost of building production ML systems drops by half every six months. This is not slowing down.