New AI Compression Shrinks Models to 2 Bits—But Is It Too Good to Be True?

A research group called ISTA has released a compressed version of the DeepSeek-V4.1-Flash AI model. They used a technique called GSQ-RCO to shrink the model down to just 2 bits per weight—think of it as squeezing a huge file into a tiny one. They claim this compressed version performs just as well as the original, which is surprising because usually compression degrades quality.

If this is true, it's a big deal. Smaller AI models can run on phones, laptops, and other devices without needing expensive cloud servers. That could make advanced AI cheaper and more accessible to everyone. You might get faster responses, lower costs, and more privacy since your data stays on your device.

But there's a catch. The benchmarks (standard tests) they used are saturated, meaning almost everyone scores near 100%. For example, on a math test called AIME25, every version scored 100. So these tests can't show if the compressed model is actually worse. Experts are asking for more sensitive measures like KLD (a way to compare how similar two models are) or PPL (perplexity, a measure of how well a model predicts text). Without those, it's hard to know if the compression really preserves quality.

So while the claim is exciting, it's not yet proven. If you're considering using this model, be cautious. It might work great for simple tasks, but for complex ones, the compression could cause subtle errors. The AI community is waiting for independent tests to verify the claims. Until then, take the 'too good to be true' hype with a grain of salt.