DeepSeek-V4-Flash-0731: 304B params, best value-per-intelligence model
Source: https://simonwillison.net/2026/Jul/31/deepseek-v4-flash-0731/
Author: Simon Willison
Date: 2026-07-31
Summary (Chinese)
DeepSeek V4 系列新成员，304B 参数（HF 上 167GB），Artificial Analysis 把它排在 MiniMax M3（428B）前面。定价 $0.14/百万输入、$0.27/百万输出，在同智能指数下成本仅为同类模型的十分之一。默认 reasoning 级别画出有问题的鹞鹕自行车，reasoning_effort 调到 high 后结果好很多。
Summary (English)
DeepSeek V4 family new member, 304B params (167GB on HuggingFace). Artificial Analysis ranks it ahead of MiniMax M3 (428B). Pricing $0.14/M input, $0.27/M output makes it potentially the best value-per-intelligence model. Default reasoning level produced a flawed pelican bicycle; high reasoning effort gave much better results.
One-liner
同智能指数下成本为同类模型十分之一，帕累托前沿左上角
*Imported from Obsidian digest notes on 2026-08-01. Content grounded in fetched source metadata via opencli.*