@MaxForAI: You'd be hard-pressed to find a better eval resource library. If you're interested in eval, these are what you should read. Thanks to @xdotli for sharing.

It’s hard to find a better eval resource library than this one.

If you’re interested in eval, these are what you should read.

Thanks to @xdotli for sharing

Xiangyi Li (@xdotli): sharing my personal library on evals 1/n

i put together the highest quality blogs, podcasts, papers, and projects on evals. additions are welcome!

The Unsloth team is terrifying.

They took China’s top open-source model, GLM 5.2, and optimized it using an extreme technique called 1-bit compression, then converted it into a lightweight GGUF format.

This means you can run GLM 5.2 entirely locally (on a 256GB Mac Studio) with no internet connection or external server needed, and at an impressive speed of 21 tokens per second (human speech is about 10-20 tokens per second).

And they didn’t stop there — they also livestreamed a showdown, pitting this compressed local model against the world’s most powerful and expensive paid cloud models: Claude 4.8 Opus and GPT-5.5.

What shocked developers in the comments the most was that this local model actually traded blows with multi-billion-dollar server clusters, delivering intelligent and precise answers that rivaled these closed-source giants.

The real winner today isn’t a specific model — it’s the concept of local inference:

From now on, your data stays 100% safe on your device
Your API bill is zero dollars, with intelligence on par with the best US companies

What a crazy but revolutionary approach.