1
00:00:00,000 --> 00:00:01,640
仅凭模型后训练的优化

2
00:00:01,640 --> 00:00:04,160
模型性能就能暴涨超过30%

3
00:00:04,160 --> 00:00:07,240
一个推理激活的参数量只有13B的模型

4
00:00:07,240 --> 00:00:09,980
居然就逼近了Cloud Open 4.8模型的性能

5
00:00:09,980 --> 00:00:12,420
这就是Deepseek V4 Flash正式版

6
00:00:12,420 --> 00:00:13,380
时隔三个月

7
00:00:13,380 --> 00:00:16,400
Deepseek V4的正式版模型终于陆续登场

8
00:00:16,400 --> 00:00:18,060
就在7月31号下午

9
00:00:18,060 --> 00:00:20,380
V4 Flash的正式版率先亮相

10
00:00:20,380 --> 00:00:22,180
相比三个月前的阅览版

11
00:00:22,180 --> 00:00:24,780
正式版Flash模型保持的架构和模型参数不变

12
00:00:24,780 --> 00:00:27,060
仍然是284B总参数量

13
00:00:27,060 --> 00:00:29,160
13B激活参数的MOE模型

14
00:00:29,160 --> 00:00:32,020
但正式版重新进行了模型的后训练

15
00:00:32,020 --> 00:00:35,400
不仅Agent和编程的性能实现了跨越式的增长

16
00:00:35,400 --> 00:00:38,360
而且模型的响应格式也进行了全面升级

17
00:00:38,360 --> 00:00:40,940
在原先的OpenEye Compilation API的基础上

18
00:00:40,940 --> 00:00:43,220
进一步适配了OpenEye的Responses API

19
00:00:43,220 --> 00:00:46,060
并且官宣正式全面接入Codex的生态

20
00:00:46,060 --> 00:00:50,940
本期外科为你深度解读Deepseek V4 Flash模型特性和性能评测结果

21
00:00:50,940 --> 00:00:53,280
并为大家详细介绍V4模型

22
00:00:53,280 --> 00:00:55,380
接入全新的Parnest Agent API

23
00:00:55,380 --> 00:00:57,540
Responses API的使用方法

24
00:00:57,540 --> 00:00:59,440
以及接入Codex的完整流程

25
00:00:59,440 --> 00:01:03,380
帮助大家离门槛上手使用全新的Deepseek V4正式版模型

26
00:01:03,380 --> 00:01:05,260
同时国外科的文字版资料

27
00:01:05,260 --> 00:01:07,500
OpenEye Responses API的使用方法

28
00:01:07,500 --> 00:01:09,800
以及Codex接入的文字版教程

29
00:01:09,800 --> 00:01:11,940
都已经上限至复犯AI社区

30
00:01:11,940 --> 00:01:14,460
大家扫描频道的二维码即可免费领取

31
00:01:14,460 --> 00:01:16,220
首先是性能方面

32
00:01:16,220 --> 00:01:18,440
根据官方发布的评测结果

33
00:01:18,440 --> 00:01:21,620
Deepseek V4 Flash已经全面超越了GM5.2

34
00:01:21,620 --> 00:01:23,440
和Deepseek V4 Pro的预览版

35
00:01:23,440 --> 00:01:27,280
甚至在一些评分指标上已经逼近了OP4.8模型

36
00:01:27,280 --> 00:01:29,800
而根据实际的测试结果来看

37
00:01:29,800 --> 00:01:31,320
模型的阉城性能

38
00:01:31,320 --> 00:01:32,200
Agent性能

39
00:01:32,200 --> 00:01:35,860
以及对真实物理世界的理解能力都进步巨大

40
00:01:35,860 --> 00:01:38,060
甚至困扰Deepseek已久的

41
00:01:38,060 --> 00:01:40,680
鸭子骑自营车的SVG动画生成

42
00:01:40,680 --> 00:01:43,080
都被V4 Flash正式版轻松拿下

43
00:01:43,080 --> 00:01:45,500
简直和此前的V4预览版

44
00:01:45,500 --> 00:01:46,820
判若两个模型

45
00:01:46,820 --> 00:01:50,020
当然如果说性能进步是常规迭代

46
00:01:50,020 --> 00:01:54,760
那么本次V4正式版模型关心拥抱OpenAI的Responses API技术体系

47
00:01:54,760 --> 00:01:58,060
则预示着正式版模型将会是全面接容OpenAI

48
00:01:58,060 --> 00:02:01,540
Codex技术生态的原生Agent基座模型

49
00:02:01,540 --> 00:02:02,980
而这对于开发者来说

50
00:02:02,980 --> 00:02:06,140
无疑是一次巨大的第一层API开发范式的转变

51
00:02:06,140 --> 00:02:10,220
同时也预示着跋越即将发布的Deepseek Harness Agent

52
00:02:10,220 --> 00:02:13,320
将会和OpenAI的基础通信范式保持一致

53
00:02:13,320 --> 00:02:15,820
那么接下来就让我给大家简单解释下

54
00:02:15,820 --> 00:02:17,920
到底什么是Responses API

55
00:02:17,920 --> 00:02:19,340
以及其基础使用方法

56
00:02:19,340 --> 00:02:23,900
然后再详细介绍如何将Deepseek V4正式版模型接入Kodaks

57
00:02:23,900 --> 00:02:25,280
在过去很长长时间

58
00:02:25,280 --> 00:02:28,240
OpenAI制定的Chat Compilations通信标准

59
00:02:28,240 --> 00:02:31,400
是唯一跨模型统一的模型底层通信规范

60
00:02:31,400 --> 00:02:35,120
那么哪怕是前段时间刚上线不久的Kimi K3模型

61
00:02:35,120 --> 00:02:39,060
底层仍然采用的是Chat Compilations作为基础模型响应格式

62
00:02:39,060 --> 00:02:42,920
这其实是一种面向消息对列的高效编辑的一类API

63
00:02:42,920 --> 00:02:46,160
开发者可以非常高效地完成Messages的列表编辑

64
00:02:46,160 --> 00:02:49,600
并通过修改消息列表来引导模型完成任务

65
00:02:49,600 --> 00:02:52,040
但或许是因为消息对列的编辑

66
00:02:52,040 --> 00:02:55,100
只适合打造聊天机器人而不适合打造agent

67
00:02:55,100 --> 00:02:56,560
于是去年3月11号

68
00:02:56,560 --> 00:03:01,040
OpenAI推出了全新的全面适配agent任务的模型调用

69
00:03:01,040 --> 00:03:02,680
新版是Responses API

70
00:03:02,680 --> 00:03:05,400
而我当时也在第一时间对其进行了解读和评测

71
00:03:05,400 --> 00:03:09,480
不同于Chat Compilations API是面向对话来构建应用

72
00:03:09,480 --> 00:03:14,280
Responses API是一种面向任务来构建应用的更高级的API

73
00:03:14,280 --> 00:03:16,980
用户只需要描述任务关联工具

74
00:03:16,980 --> 00:03:20,920
说的题词词就能自动构建一个agent loop循环非常高效

75
00:03:20,920 --> 00:03:25,080
或者你也可以将其理解成Responses API就是OpenAI版的LongChain

76
00:03:25,080 --> 00:03:30,220
而现在DeepC V4正式版模型也是国内首个全面接入Responses API的模型

77
00:03:30,220 --> 00:03:34,040
而我们在第一时间对其整体性能表现进行了深度评测

78
00:03:34,040 --> 00:03:37,920
在海浪工具Skills和MCP调度情况下运行非常稳定

79
00:03:37,920 --> 00:03:42,500
同时我们在第一时间制作了DeepC Responses API的接入指南和使用教程

80
00:03:42,500 --> 00:03:45,520
并上线至复犯AI社区大扫码即可免费领取

81
00:03:45,520 --> 00:03:47,920
当然对于DeepC V4模型来说

82
00:03:47,920 --> 00:03:52,840
全面拥抱Responses API体系也就意味着可以不贫接入Codex的完整功能生态

83
00:03:52,840 --> 00:03:57,080
而官方给出的V4正式版接入Codex的流程也非常简单

84
00:03:57,080 --> 00:04:02,720
只需要一行命令即可同步切换Codex Cli,Codex App和ChainGBT的基座模型

85
00:04:02,720 --> 00:04:05,020
将其切换为DeepC V4正式版模型

86
00:04:05,020 --> 00:04:08,920
同时完成原始配置的备份并且可以随时复原

87
00:04:08,920 --> 00:04:10,360
以Windows系统为例

88
00:04:10,360 --> 00:04:12,500
完整切换基座模型的流程如下

89
00:04:12,500 --> 00:04:14,260
首先先打开PowerShell

90
00:04:14,260 --> 00:04:16,540
然后输入IRM命令

91
00:04:16,540 --> 00:04:20,880
从DeepC官方下载Codex配置文件修改脚本并自动运行

92
00:04:20,880 --> 00:04:25,240
然后根据提示进行模型选取或者恢复原始配置

93
00:04:25,240 --> 00:04:29,540
配置模型可以根据实际情况选择V4 Flash或者Pro正式版模型

94
00:04:29,540 --> 00:04:31,720
并根据提示设置API Key

95
00:04:31,720 --> 00:04:32,640
设置完成后

96
00:04:32,640 --> 00:04:35,420
只要看到Installation Complete的提示

97
00:04:35,420 --> 00:04:36,880
就说明配置成功

98
00:04:36,880 --> 00:04:40,860
而如果之后要切换模型或者恢复Codex的原始配置

99
00:04:40,860 --> 00:04:44,720
则再次运行相同脚本并进行相应选择即可

100
00:04:44,720 --> 00:04:47,700
然后打开HitGPT或者Codex应用

101
00:04:47,700 --> 00:04:49,500
看到左下方是DeepSeq

102
00:04:49,500 --> 00:04:51,260
并且模型是自定义

103
00:04:51,260 --> 00:04:52,960
就说明已经配置成功了

104
00:04:52,960 --> 00:04:56,400
接下来即可进行任意对话或者执行各类任务

105
00:04:56,400 --> 00:04:57,500
整个过程非常简单

106
00:04:57,500 --> 00:05:00,100
并且Codex Click也会同步发生修改

107
00:05:00,100 --> 00:05:02,700
而紧接着我们就在Codex命令行中

108
00:05:02,700 --> 00:05:05,440
对V4 Flash正式版模型进行一轮深度评测

109
00:05:05,440 --> 00:05:10,080
期间我们让V4 Flash围绕着一个3万多行多模态PDF检锁系统

110
00:05:10,080 --> 00:05:12,020
进行代码审查和结构优化

111
00:05:12,020 --> 00:05:15,540
一条指令即可完成整个Codex的

112
00:05:15,540 --> 00:05:18,960
Plan Task Subagent Web Search等各项功能的驱动

113
00:05:18,960 --> 00:05:22,960
并最终完成代码库的审核优化测试一整观者流程

114
00:05:22,960 --> 00:05:24,400
非常顺畅高效

115
00:05:24,400 --> 00:05:26,800
这次DeepSeq V4 Flash模型发布非常迅速

116
00:05:26,800 --> 00:05:28,800
在API全面上线同时

117
00:05:28,800 --> 00:05:31,340
模型权重也正式在Hugging Face上开源

118
00:05:31,340 --> 00:05:34,080
当然V4 Flash其实还只是开位赛

119
00:05:34,080 --> 00:05:37,840
还只是接下来V4 Pro正式版和DeepSeq版Cloud Code

120
00:05:37,840 --> 00:05:40,820
DeepSeq Harness Agent发布的前奏

121
00:05:40,820 --> 00:05:43,820
按照30%的性能涨幅预期

122
00:05:43,820 --> 00:05:44,700
我们有留相信

123
00:05:44,700 --> 00:05:47,260
V4 Pro正式版将会是全球第一款

124
00:05:47,260 --> 00:05:51,200
性能全面超越OPPO4.8的全新一代开源旗舰模型

125
00:05:51,200 --> 00:05:53,160
好了 以上就是本期视频的全部内容

126
00:05:53,160 --> 00:05:54,740
感谢大家的关注和三连支持

127
00:05:54,740 --> 00:05:56,840
更多DeepSeq V4正式版的学习资料

128
00:05:56,840 --> 00:05:59,700
技术文档 模型部署跟Agent开发教程等

129
00:05:59,700 --> 00:06:01,740
都已经上限至复办AI技术社区

130
00:06:01,740 --> 00:06:03,100
大家扫码即可领取

131
00:06:03,100 --> 00:06:06,680
五十九天 专注大家提供最扎实优质的技术内容

132
00:06:06,680 --> 00:06:08,280
我们下个视频 再见