Efficiency and local models are beating the bigger-is-better story

Llama-modchip claims to run a 284B DeepSeek on 64GB of RAM with no GPU. Commenters immediately asked about quantization and guessed it is 2-bit, so quality is the open question. A $10 OpenStick homelab post leans on the line "given how expensive RAM is these days." The MSR thread on the end of Microsoft Research's AI Frontiers group has a top comment saying the US over-invested in frontier hardware while the next breakthroughs will likely be radical optimizations. A quadrillion-parameter paper sits at the other end of the spectrum.

In the older Ask HN on replacing Claude/GPT with a local model (48542100), people report that Qwen 3.6 27B dense is roughly Haiku 4.5 level, and that Gemma 4 on an M4 is too slow. One person runs DeepSeek V4 Flash on two RTX Pro 6000 cards at 160 tok/s.

The through-line is that people want cheaper, owned inference and are willing to trade quality for it. The unresolved problem is that quality and speed at consumer scale are still uneven, and the claims are hard to verify.