320B parameters, but only 15B actually fire at a time. That is the part of IQu… · @harden_369741 · TwKit

@harden_369741

320B parameters, but only 15B actually fire at a time. That is the part of IQuest-Q1 I keep thinking about. It is built for agentic CLI work, not chat demos. The recipe is three stages: stable pre-training, fine-grained mid-training, then multi-scaffold post-training with SFT, domain RL and model merging. For me, the bet is simple: a big sparse model that stays cheap to run could make terminal agents practical for everyday coding, not just flagship demos. Would you trust a 15B-active model with your repo, or still reach for the biggest dense model first? Paper: https://arxiv.org/abs/2610.30949 #AI #CodingAgents #OpenSource #LLM

曝光 8 · 评论 0 · 点赞 0 · 书签 0