大模型
pub.towardsai.net ·
@Chew Loong Nian - AI ENGINEER
NVIDIA 4-bit Qwen 3.8 27B 为何下载体积是 21.9 GB 而非 13.9 GB？
文章指出 NVIDIA 的 NVFP4 版 Qwen 3.8 27B 权重实际为 21.9 GB，而非按“278 亿参数 × 4 bit”估算的 13.9 GB，原因是该版本采用混合精度（约 5.5 bits，仅约 66% 权重为 4-bit）。作者称可通过读取 87 KB 的 config.json，用约 110 行脚本在下载前精确预测张量字节数（21,921,428,072），并说明额外约 8 GB 的来源及影响该数值的设置项。文中提到该模型于 9 月 4 日上传 Hugging Face，9 月 27 日页面显示上月下载量 299,067 次。
AI推荐理由 文章解释 NVIDIA NVFP4 版 Qwen 3.8 27B 权重下载体积为 21.9 GB 而非按 4-bit 估算的 13.9 GB，原因是混合精度（约 5.5 bits，仅 66% 权重为 4-bit），并称可用 110 行脚本读取 config.json 精确预测张量字节数。提供了具体字节数（21,921,428,072）和 Hugging Face 下载量（299,067）等可核对细节，对量化部署有参考价值；但正文片段仅到“额外 8 GB 去向”处截断，脚本与设置细节未展示，且为 Medium 会员文章，无法独立验证其结论。
62
正文
Member-only story
Nvidia
Artificial Intelligence
Machine Learning
Why Is NVIDIA's 4-Bit Qwen 3.8 27B a 21.9 GB Download, Not 13.9?
Chew Loong Nian - AI ENGINEER
8 min read
·
7 hours ago
--
Listen
Share
Press enter or click to view image in full size
Only 66% of its weights are 4-bit. A 110-line script reads config.json and predicts the checkpoint’s tensor bytes exactly (21,921,428,072) before you download it, then shows which setting moves the number.
NVIDIA’s NVFP4 build of Qwen 3.8 27B is 21.9 GB of weights, not the 13.9 GB that “27.8 billion parameters at 4 bits” suggests, and you can predict that to the byte from its 87 KB config.json before you download the weights.
Below: where the extra 8 GB goes, the script, and the setting that changes it.
01 — The number on the download page
nvidia/Qwen3.8-27B-NVFP4 went up on Hugging Face on September 4, in a commit whose title includes "AutoQuant mixed-precision NVFP4/FP8, ~5.5 bits". On September 27 its page showed 299,067 downloads last month.
Its three shards total 21,921,697,280 bytes; the index says the tensors take 21,921,428,072 of them, and the rest is shard headers. The naive guess, 27.78 billion parameters × 4 bits, is 13.89 GB. The real download is 8 GB heavier.
02 — Where the other 8 GB goes