DeepSeek v4.1 Flash 1017 – Why the “million‑token” LLM finally feels affordable The Lead “A 1 M‑token context used to cost a data‑center the price of a small‑car loan.” When DeepSeek unveiled v4.1‑Flash 1017 on September 17, 2026, the headline sparked more than a few skeptical comments on Hacker News. The claim sounded almost mythical: a 284‑billion‑parameter transformer that can remember an entire novel‑length document and fit into the memory budget of a single A100 ...