TurboQuant Model nears lossless 4-bit weights

This is one of the more credible "quantize everything" experiments in a while: the repo is not just shaving bits, it's showing that a residual pass can recover most of the quality loss. The caveat is that the win depends on a fairly sophisticated kernel path, so the real question is how much of the headline survives outside the authors' benchmark setup.

• –On Qwen3.5-0.8B, 4+4 residual gets 14.28 PPL vs 14.29 bf16, which is close enough to feel operationally meaningful.
• –Plain 4-bit is still a useful memory play, but it pays a real accuracy tax, so the residual stage is doing most of the heavy lifting.
• –The 4B edit is interesting because 4+2 residual slightly beats bf16 on PPL while 4+4 keeps KLD much lower, which is a good reminder that perplexity alone doesn't tell the whole story.
• –The implementation story matters: on-the-fly dequantization plus fused CuTile/Triton kernels is what keeps this from becoming an academic demo that falls apart in production.
• –There is already some community debate about TurboQuant's theoretical lineage, so I'd treat the "near-optimal" claim as promising but still worth validating in your own stack.