开启左侧

DeepSeek-V3.1开源发布

[复制链接]
在线会员 IMeYQk 发表于 2025-8-22 20:45:55 | 显示全部楼层 |阅读模式 打印 上一主题 下一主题 |快速收录
DeepSeek-V3.1 是一个撑持思考情势战非思考情势的混淆模子。取以前的版底细比,此次升级正在多个圆里戴去了改良:

    混淆思考情势:颠末变动谈天模板,一个模子能够共时撑持思考情势战非思考情势。

    更智能的东西挪用:颠末后锻炼劣化,模子正在东西使用战代办署理任务圆里的功用清楚进步。

    更下的思考服从:DeepSeek-V3.1-Think 正在答复品质上取 DeepSeek-R1-0528 相称,但是照应速率更快。

DeepSeek-V3.1 是正在 DeepSeek-V3.1-Base 的根底上截至后锻炼获得的,后者颠末二阶段少高低文扩大办法鉴于本初 V3根底 查抄面建立,依照了本初 DeepSeek-V3 陈述中概括的办法。咱们颠末汇集分外的少文档并年夜幅扩大二个锻炼阶段去扩大了咱们的数据散。32K 扩大阶段已经增加了 10 倍至 630B tokens,而 128K 扩大阶段则增加了 3.3 倍至 209B tokens。别的,DeepSeek-V3.1 使用 UE8M0 FP8 格局的数据截至锻炼,以保证取微缩搁数据格局兼容。
模子下载

模子总参数数激活参数数高低文少度下载
DeepSeek-V3.1-Base671B37B128KHuggingFace | ModelScope
DeepSeek-V3.1671B37B128KHuggingFace | ModelScope
谈天模板


咱们的谈天模板概略睹 tokenizer_config.json 战 assets/chat_template.jinja。那里是一个扼要描绘。
非思考情势

第一回开


前缀: <|begin▁of▁sentence|>{体系提醒}<|User|>{盘问}<|Assistant|> </think>

给定此前缀,DeepSeek V3.1 正在非思考情势下天生对于盘问的照应。取 DeepSeek V3 差别,它引进了一个分外的标识表记标帜 </think>。
多回开


高低文: <|begin▁of▁sentence|>{体系提醒}<|User|>{盘问}<|Assistant|> </think>{照应}<|end▁of▁sentence|>...<|User|>{盘问}<|Assistant|> </think>{照应}<|end▁of▁sentence|>

前缀: <|User|>{盘问}<|Assistant|> </think>

颠末将高低文战前缀跟尾起去,咱们得到了准确的盘问提醒。
思考情势

第一回开


前缀: <|begin▁of▁sentence|>{体系提醒}<|User|>{盘问}<|Assistant|><think>

思考情势的前缀类似于 DeepSeek-R1。
多回开


高低文: <|begin▁of▁sentence|>{体系提醒}<|User|>{盘问}<|Assistant|> </think>{照应}<|end▁of▁sentence|>...<|User|>{盘问}<|Assistant|> </think>{照应}<|end▁of▁sentence|>

前缀: <|User|>{盘问}<|Assistant|><think>

多回开模板取非思考情势下的多回开谈天模板差异。那表示着最初一轮中的思考标识表记标帜将被简略,但是正在每一轮高低文中皆保存 </think>。
东西挪用


东西挪用撑持正在非思考情势下使用。格局为:

<|begin▁of▁sentence|>{体系提醒}\n\n{东西描绘}<|User|>{盘问}<|Assistant|> </think> 此中东西描绘是
## Tools
You have access to the following tools:

### {tool_name1}
Description: {description}

Parameters: {json.dumps(parameters)}

IMPORTANT: ALWAYS adhere to this exact format for tool use:
<|tool▁calls▁begin|><|tool▁call▁begin|>tool_call_name<|tool▁sep|>tool_call_arguments<|tool▁call▁end|>{additional_tool_calls}<|tool▁calls▁end|>

Where:
- `tool_call_name` must be an exact match to one of the available tools
- `tool_call_arguments` must be valid JSON that strictly follows the tool's Parameters Schema
- For multiple tool calls, chain them directly without separators or spaces
代码代办署理


咱们撑持多种代码代办署理框架。请参照上述东西挪用格局去创立您自己的代码代办署理。一个示例睹 assets/code_agent_trajectory.html。
搜刮代办署理


咱们为正在思考情势下搜刮东西挪用设想了一定的格局,以撑持搜刮代办署理。

关于需要会见内部或者最新疑息的庞大成就,DeepSeek-V3.1 能够颠末多轮东西挪用历程使用用户供给的搜刮东西。

请参阅 assets/search_tool_trajectory.html 战 assets/search_python_tool_trajectory.html理解 具体的模板。
评介

种别基准(目标)DeepSeek V3.1-非思考DeepSeek V3 0324DeepSeek V3.1-思考DeepSeek R1 0528
通用





MMLU-Redux (EM)91.890.593.793.4

MMLU-Pro (EM)83.781.284.885.0

GPQA-Diamond (Pass@1)74.968.480.181.0

人类最初的测验 (Pass@1)--15.917.7
搜刮代办署理





BrowseComp--30.08.9

BrowseComp_zh--49.235.7

人类最初的测验 (Python + 搜刮)--29.824.8

SimpleQA--93.492.3
代码





LiveCodeBench (2408-2505) (Pass@1)56.443.074.873.3

Codeforces-Div1 (评分)--20911930

Aider-Polyglot (精确率)68.455.176.371.6
代码代办署理





SWE 考证 (代办署理情势)66.045.4-44.6

SWE-bench 多语言 (代办署理情势)54.529.3-30.5

Terminal-bench (Terminus 1 框架)31.313.3-5.7
数教





AIME 2024 (Pass@1)66.359.493.191.4

AIME 2025 (Pass@1)49.851.388.487.5

HMMT 2025 (Pass@1)33.529.284.279.4

正文:
    搜刮代办署理使用咱们的内部搜刮框架截至评介,该框架使用贸易搜刮 API + 网页过滤器 + 128K 高低文窗心。R1-0528 的搜刮代办署理成果是颠末预约义的事情流截至评介的。SWE-bench 使用咱们的内部代码代办署理框架截至评介。HLE 使用仅文簿本散截至评介。
使用示例


import transformers

tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.1")

messages = [
    {"role": "system", "content": "You are a helpful assistant"},
    {"role": "user", "content": "Who are you?"},
    {"role": "assistant", "content": "<think>H妹妹</think>I am DeepSeek"},
    {"role": "user", "content": "1+1=?"}
]

tokenizer.apply_chat_template(messages, tokenize=False, thinking=True, add_generation_prompt=True)
# '<|begin▁of▁sentence|>You are a helpful assistant<|User|>Who are you?<|Assistant|></think>I am DeepSeek<|end▁of▁sentence|><|User|>1+1=?<|Assistant|><think>'

tokenizer.apply_chat_template(messages, tokenize=False, thinking=False, add_generation_prompt=True)
# '<|begin▁of▁sentence|>You are a helpful assistant<|User|>Who are you?<|Assistant|></think>I am DeepSeek<|end▁of▁sentence|><|User|>1+1=?<|Assistant|></think>'
您需要登录后才可以回帖 登录 | 立即注册 qq_login

本版积分规则

发布主题
阅读排行更多+
用专业创造成效
400-778-7781
周一至周五 9:00-18:00
意见反馈:server@mailiao.group
紧急联系:181-67184787
ftqrcode

扫一扫关注我们

Powered by 职贝云数A新零售门户 X3.5© 2004-2025 职贝云数 Inc.( 蜀ICP备2024104722号 )