上手实测：Gemini 4 Argon 稀疏 MoE、单次输出 100 万 token
作者称 Argon 是原生多模态稀疏 MoE，单次可输出 100 万 token，并对比 Opus 5.5、GPT-6 Astra、Grok 4.7 的短板。
I tried Gemini 4 Argon, and it's genuinely great, so I dug in to see what's actually under the hood. It's a new Sparse MoE model that natively handles text, tables, audio, and video together and can generate up to 1M tokens in a single response without cutting off mid-thought. Last night I fed it a pile of raw PDF reports, table screenshots, and messy notes, and asked it to pull everything into one clean document. Here's where the difference showed up in real work: Opus 5.5 forces you to split big tasks into chunks because of its 128K output cap, asks endless follow-up questions, and doesn't take video or audio. GPT 6 Astra thinks way too long before starting and likes to make up facts when something is missing, hallucinating 3x more often. Grok 4.7 runs slow, adds tons of fluff, and loses details across large files. Meanwhile, Argon delivered a clean, finished result on the first try with zero nagging or canned templates. What won me over most is how effortless this AI feels. You don't have to wrestle with prompts or beg the model to finish the job, because it just gets what you mean right away. I'm definitely going to keep testing it on my daily workflows, since right now it feels like the smoothest release in a long time. Anyone else had a chance to play around with it yet, and what do you think?