After MoE and MTP, local dense models haven't seen architectural speedups
Express_Quail_1493 · reddit · 2026-10-22
A discussion asking when local dense models will get their next speed breakthrough. Most recent gains come from MoE (cutting active vs total params), at the cost of stability — a dense 27B at low total params is more reliable. Since MTP, dense models have seen little architectural attention beyond community finetunes trimming reasoning tokens. With Qwen4-27B coming, is a dense speedup breakthrough likely?
More from Infra
- Thorsten Ball demos spinning up a new orb provider on an exe.dev VM with Docker — davidcrawshaw · 2026-10-23
- Google advances Coral NPU with open architecture, compiler tools, and commercial chips for on-device AI — GoogleAI · 2026-10-23
- Weaviate cut vector index memory from 9.9 GB to 1.4 GB over three years of quantization — CShorten30 · 2026-10-23
- Rescale built an AI triage agent in 5 days with Strands Agents and Claude on Bedrock — AWS ML Blog · 2026-10-23
- Tesla and NVIDIA back $2.8B expansion of Samsung's advanced chip substrate supply — XFreeze · 2026-10-23
- Analyst Ben Bajarin: think of data centers as prime real estate — BenBajarin · 2026-10-23