JuneoAI/Laplace-1-Vector
Text Generation • 4B • Updated • 591 • 41
AI & ML interests
Trusted collaboration for people and AI. Building Laplace and Juneo.
Recent Activity
Organization Card
Juneo Lab
English
JUNEO LAB · LAPLACE-1
The trust layer for
people and agents
Laplace foresees the consequence of an action before an agent takes it, and makes the four decisions organizations delegate: who does the work, who sees the information, what runs, and what is accepted
Looking back from the age of intelligence
Every leap of civilization
was an expansion of trust
The scale of cooperation has never been set by how intelligent individuals are, but by how far strangers can trust one another
~70,000 years ago
The tribe
Trust came from
Familiarity
~5,000 years ago
Cities and empires
Trust came from
Writing and law
1602
The joint-stock company
Trust came from
Contract
1990s
The internet
Trust came from
Platforms and payments
2030s
People and agents
Trust came from
Foresight of consequences
The consensus
In the age of AI, what is scarce is stronger intelligence
Our view
Intelligence is no longer scarce
What is scarce is trust
The fifth leap
Trust will come from foreseeing consequences
Laplace builds that trust for the second workforce
Four decisions
Organizations delegate
more than actions
Whenever people and agents work together, four questions come up
Laplace foresees the consequence of each candidate, then decides
01
Allocate work
Who should own this task?
02
Share information
Who may see what, and how much?
03
Act
Run it, adjust it, or ask first?
04
Evaluate outcomes
Does the result meet the standard?
Laplace-1 · Open weights
First on most
decision benchmarks
A flagship and an on-device size, evaluated against models from OpenAI, Perplexity, ByteDance, Alibaba, and others. Complete weights, Apache-2.0
27B · Flagship
Laplace-1-Atlas
First on 7 of 9 decision benchmarks
89.4% on TrustDecision, 5.4 points ahead of the next-best model
Model card ↗4B · On-device
Laplace-1-Vector
First on 7 of 8 decision benchmarks
Ahead of IBM’s Granite Switch 4.1 8B, twice its size, on all 8. Judgment runs where agents act
Model card ↗First places include ties
Decision Forge · Open source
The standard of judgment
lives inside the organization
Public data records what happened, never what an action caused. Decision Forge turns an organization’s own policies, approvals, and tickets into verified decision data, and every label traces back to a rule and its source
xDecision context
The state, permissions, and options at the moment an agent acts
ℓJudgment standard
What this organization allows, forbids, or escalates
s′Execution receipt
What actually changed after the action
The better Laplace judges, the more work organizations delegate
More delegation brings more receipts, and more receipts teach Laplace the organization
Juneo Desktop
A local-first agent for macOS. Your files stay on your Mac, and every step leaves a record you can check
juneo.ai ↗Let your organization
trust agents with real work
Juneo Lab · The trust layer for people and agents
Era images: Lascaux cave paintings; Thomas Cole, The Course of Empire: The Consummation (1836); H. C. Vroom, Return to Amsterdam of the Second Expedition to the East Indies (1599); NASA, Earth at Night. Public domain, via Wikimedia Commons
中文
JUNEO LAB · LAPLACE-1
智能时代的
可信协作基础设施
Laplace 在 Agent 动手之前推演后果，替组织做好四类判断：工作交给谁、信息给谁、动作做不做、结果收不收
站在智能时代的终局回望
人类文明的每一次跃迁
都是一次信任的扩张
协作的规模，从来不取决于个体有多聪明，而取决于陌生人之间能否互相信任
约 7 万年前
部落
信任来自
熟人
约 5000 年前
城邦与帝国
信任来自
文字与法律
1602 年
股份公司
信任来自
契约
1990 年代
互联网
信任来自
平台与支付
2030 年代
人与 Agent 的组织
信任来自
对后果的预见
共识
AI 时代，稀缺的是更强的智能
我们的判断
智能已不稀缺
稀缺的是信任
第五次跃迁
信任来自对后果的预见
Laplace 为第二种劳动力建造信任基础设施
四类判断
组织交给 AI 的
从来不只是一个动作
人与 Agent 一起工作时，每天都要回答四个问题
Laplace 先推演每个候选的后果，再给出判断
01
分配任务
这件事交给谁？
02
共享信息
这条信息能给谁、给多少？
03
执行操作
直接做、调整，还是先请示？
04
验收成果
结果合不合格？
Laplace-1 · 开源权重
同台 OpenAI、字节、阿里
拿下多数第一
旗舰与端侧两个版本，权重完整开源，Apache-2.0 协议
4B · 端侧
Laplace-1-Vector
8 项决策基准 · 7 项第一
全部 8 项超过参数量是它两倍的 IBM Granite Switch 4.1 8B，判断在 Agent 身边完成
查看模型卡 ↗“第一”含并列
Decision Forge · 开源
判断标准
只存在于组织内部
公开数据只记录发生了什么，不记录一个动作造成了什么。Decision Forge 把组织自己的制度、审批和工单变成可验证的决策数据，每个标签都能追溯到一条规则，每条规则都能追溯到原文
x决策上下文
Agent 动手那一刻，组织的状态、权限与候选项
ℓ判断标准
这个组织允许什么、禁止什么、什么需要升级
s′执行回执
动作完成之后，组织真实发生了什么
判得越准，组织交出的工作越多
交出的越多，回执越多，Laplace 就越懂这个组织
让你的组织
敢把工作交给 Agent
Juneo Lab · 智能时代的可信协作基础设施
时代图片：拉斯科洞穴壁画；Thomas Cole《帝国的历程·鼎盛》(1836)；H. C. Vroom《第二次东印度远航返回阿姆斯特丹》(1599)；NASA 地球夜景。均为公有领域，来自 Wikimedia Commons
datasets 0
None public yet