IQuest-Q1 at 320B: weights 650GB, 8 GPUs — license and tokenizer checked

The piece is deliberately framed as a partial-verification exercise. Because the model's size put live inference out of reach — BF16 weights at roughly 650GB and an official deployment recipe assuming 8 GPUs via vLLM/SGLang — the author drew an explicit line: nothing that was not actually run would be described as run. That stance shapes what the piece does and does not cover.

What remained verifiable was checked at the file level. The benchmark table is quoted from the paper rather than reproduced, and the author states directly that those are the authors' own reported figures. The release's config.json was confirmed to match the architecture description, and the tokenizer was downloaded and exercised on sample text.

One minor discrepancy surfaced along the way: the tokenizer vocabulary still contains image and video placeholder tokens, even though the model is documented as text-input only. The author calls this a residual rather than a practical problem, but notes it as a small gap between how the model is described and what the vocabulary actually contains.