2026-11-05 · America/Los_Angeles · 社区动态 · #19

Rather Qwen4-27b at 1M ctx than a big MOE at 128k

Larger context just gets more work done rather than having to re-read files ETC id much rather use my extra resources on context window. Instead of more params. i get my 400k ctx qwen3.8-27b + web search search get the latest up-to-date docs for its lack of world knowledge and i don’t waste extra Vram storing data that i dont need. Also more active params makes the llm to hold a wider mental-model of the task. Qwen4-27b will be architected to make 1M more reliable and using it for context on a smaller dense model seems more productive for me. Anyone. Else notice agentic session really shines…

热度 37.5 / 100；排名与评分保留该期记录。