# DeepSeek V4 Flash 0731 纯后训练提升 10 分

- 来源：Chubby♨️ (@kimmonismus)
- 发布时间：2026-07-31 20:06
- AIHOT 分数：43
- AIHOT 链接：https://aihot.virxact.com/items/cms8x4jpt04w1ro7vz2mdrn71
- 原文链接：https://x.com/kimmonismus/status/2083162331060605315

## AI 摘要

DeepSeek V4 Flash 0731 在 Artificial Analysis Intelligence Index 上得分 50，较前代提升 10 分，超越 49B 活跃参数的 Pro 预览版，且智能体任务 token 用量减少 12%。

## 正文

Same architecture， same 13B active parameters， same price， and a 10-point jump from post-training alone！

DeepSeek V4 Flash 0731 now beats the 49B-active Pro preview on agentic work while using 12% fewer tokens.

DeepSeek found a remarkable amount of untapped capability in weights it already had.

The cache price is truly absurd. One million cached input tokens costs only $0.0028. This makes recurring agents with large， stable system prompts， codebases， or documents extremely cheap （reminds me of the DeepSeek 4 tech paper！）

DeepSeek thus delivers intelligence nearly on par with Luna at around 60% lower cost per task. However， Luna is faster， multimodal， and significantly less wasteful with tokens； following the price cut， the premium is now only about 2.4 times the cost， down from the previous 12-fold difference.

This makes me super freaking excited for the next DeepSeek 4 （4.5？） pro！

The whale is back， though todays its more a dolphin.

### 引用推文

> Artificial Analysis：DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, a 10-point jump over DeepSeek V4 Flash (released April 2026) that puts it 6 poin...