Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU

442 points · 196 comments on HN · read original →

Points and comments are a snapshot, not live.

Article body wasn't reachable. The HN discussion summary is below.

What commenters are saying

Commenters focus heavily on the quantization level, with many speculating it is a 2-bit quant. Several users report that sub-4-bit quants are often incoherent and suspect this model will not be useful for real work. A minority counter that lower quants of specific models (like DeepSeek V4 Flash) can still be valuable when that is all your hardware can run. Others point out the README is poorly written and conflates different quant sizes, making the claims confusing. A few discuss the performance trade-offs of running large models on high-RAM systems versus smaller models on GPU.