拆解 DeepSeek V4.1 Flash 词表：1249 个保留行含控制标签

stacc 'you're all unimpressive' overflow @STACCoverflow原推发布于 2026.10.24

在 X 上打开原推

I pulled DeepSeek V4.1 Flash's vocab apart to see what's in the "unused" rows. 129,280 rows, 1,249 reserved: 799 place_holder_no_N, 422 multimodal span slots, FIM/repo/tool framing, and a set of internal control tags shipped in the open weights — <query> <search> <read_url> <extracted_url> <authority> <domain> <entity> <political> <safety> <answer>. Fed as single token ids (not text), greedy, local run: – place_holder rows 0, 1, 100, 500, 798: all five answer in Chinese with "OK, send me the content / what do you need." The blank rows behave like an empty turn — no stored content, and the model's default language under them is Chinese. – <political> alone → "I'm sorry, but I can't help with that request." A refusal with zero words of input. – <authority> alone → switches to Chinese and adopts an "authoritative judgment" persona. – <safety>, <query>, <read_url>, <answer>, <begin_sys> alone → hallucinated nearby text ("syc", "Rights", "utf", "Nöen", an essay intro): live embeddings, no secret behaviour. The point: a single embedding row can carry an instruction. <political> is a refusal written into one token vector, no text attached. That's the same mechanism we use in leCore to write facts into unused rows — and DeepSeek ships 799 blank rows that could hold anything. The five I checked are empty. Sweep of all 799 next. Receipts on request.