Stop Ignoring Open Source AI

Open source AI is getting closer to the frontier faster than most people realize, and the release of GLM 5.2 by Z.ai, the company also known as Zhipu, is the clearest signal yet. In this walkthrough I break down what open source and open weights actually mean, where GLM 5.2 lands against the state of the art, and why I think this year could be the one where open models catch up.

1. Open source vs closed source (0:15)

First, the basics. Closed source and open source AI are different kinds of models you can run. Closed source means the model is only available to the parent company. Open source means it is available to any company out there.

The practical consequence is huge: if there is an open source model on the market, you can download it to your own computer, and if your machine is beefy enough, you can run it locally. That means zero data privacy issues, because nothing ever leaves your hardware. No API, no vendor, no terms of service between you and the model.

2. Open weights and data sovereignty (0:50)

On top of being open source, these models are also open weight. That means you can actually look inside the brain of the model to understand how the parameters are configured. The model stops being a black box service and becomes an asset you can inspect.

This is massive for anyone who wants sovereignty over their data, whether you are in the US or in Europe. You are not trusting a vendor's promises; you are running and inspecting the model yourself. For companies with strict data requirements, this is the difference between using AI and not using it at all.

3. Where GLM 5.2 lands on the benchmarks (1:14)

Now look at the chart I show in the video: benchmarks comparing GLM 5.2 against the state of the art, which in this case is Opus 4.8 and also Fable 5.

Yes, GLM 5.2 is still lagging behind, ranking around fourth to fifth position overall. But before you dismiss that, you need the context of what it took to get there, because that ranking is more impressive than it looks.

4. China's hardware disadvantage (1:20)

China is competing with hardware that is less performant than what the US has. The US has Nvidia. China still does not have a company as strong as Nvidia, so these models are being trained on Huawei chips, which are less performant.

Getting into the top five of the top ten models out there while training on weaker silicon is a massive feat. That is the real story behind the benchmark chart: the gap is closing even with a hardware handicap.

5. Kimi K2 and the rest of the field (1:46)

GLM is not alone. Breaking things down across other benchmarks, Chinese models like Kimi K2 are also within the top five quite consistently, rivaling Gemini, OpenAI's ChatGPT, and the models from Anthropic.

This is no longer one surprising release. It is a pattern: multiple open models sitting next to the closed frontier across benchmark after benchmark.

6. Why open source may catch up this year (2:05)

Here is where it gets exciting. Within the next couple of months, we might see open source catch up with closed source AI. That would make AI even more available to people who already have computers sitting at home, or who are willing to purchase servers to run their own instances.

Keep a close eye on what is being done in China with GLM 5.2 and Kimi K2. I am pretty sure that before the end of this year, we might start to see open source models with the exact same capabilities as Fable 5. That is a bold statement, but with the pace of advances in the open source space right now, it would not be surprising if it happens.

In my next videos I will show you how to make the most out of GLM 5.2, so this is the moment to stop ignoring the open source side of AI.

Tips

• Do not read benchmark rank in isolation. Factor in the hardware the model was trained on and the trajectory, not just today's position.
• If data sovereignty matters to your business, start evaluating open weight models now, before you need them.
• Running locally requires a beefy computer. Check your hardware honestly before you commit to a local deployment plan.

Where to go next

Models are only half the story: the application around them decides what they can actually do, and I break that down in What Claude Code really is. For a look at how the app layer is shifting on OpenAI's side, watch ChatGPT will die because of this.