Laplace-1-Vector by JuneoAI

Laplace-1-Vector by JuneoAI

Judges which action an AI agent should take, who should own a task and who may see what, as a 4B model that runs on your own hardware so organizational data stays in house.

Made by JuneoAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Laplace-1-Vector is a 4B decision model from JuneoAI, the small edge member of the Laplace-1 family. It is built to sit beside an AI agent and judge what an action would cause inside an organization before the action runs. It runs on your own hardware, so the data stays with you.

Who it suits: It suits teams that let AI agents act inside a company and want a check on those actions that does not send data outside.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Sales operations | Revenue teams | Decide who should own a new lead | Score a lead and name the people best placed to follow up
Contracts and legal | Operations staff | Route work above an approval limit | Send a large contract to its owner instead of acting alone
Field services | Dispatch teams | Let agents act on routine jobs | Release a repair order without asking a person
Professional services | Delivery teams | Check results before sign-off | Compare a deliverable with its acceptance criteria

If you run no agents or have no decisions to route, you can skip it.

Where to use it:

• Python scripts that use the transformers library
• Next to an agent, on your organization's own hardware
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Allocate work | Team leads | A ranking of who should own a task
Share information | Security and data teams | A call on who may see what, and how much
Act | Agent builders | A verdict to run, adjust or ask first
Evaluate outcomes | Reviewers | A check of a result against its standard

How to set it up: You need Python and about 8.2 GB of disk space for the BF16 weights, or about 2.1 GB for a 4-bit copy by Hugging Face's estimate.

• Python with pip
• The huggingface_hub package
• The transformers library to load the files

1. Install huggingface_hub.
2. Run the download command to save the model files on your computer.
3. Load the files with transformers in your own script.

pip install -U huggingface_hub
hf download JuneoAI/Laplace-1-Vector

How to use it: You give the model the state of the organization, the chain of authority behind the agent and the candidate actions. A decision takes one forward pass. You read a spread over your candidates from the next-token scores, with no sampling and no parsing. The card has a usage section with the exact format.

Example: An agent holds a contract above its limit and has three options: send it, adjust it or ask first. You pass the context and the three candidates. The model returns scores for each option, with the favored one highest and the strength showing how firmly the evidence points there.

For newcomers and experts: Think of it as a quick second opinion for an agent. Before the agent acts, it asks which option is best and how sure the model is. A first small test is two or three candidate actions for one task.

The model learns what the next state of an organization looks like after a candidate action, given who authorized the agent. Public text only records what happened, not what an action caused inside a company, so the card positions this as filling that gap. It is built on XHToken/Spark-X2.6-4B, and the larger 27B Laplace-1-Atlas is the flagship of the family.

Organization state and authority chain
Candidate actions go in
One forward pass
Scores for each candidate

Good to know: The authors report a top rank on 7 of the 8 benchmarks they ran, with ties counted, against models from several labs. On COMPASS, OrgAccess and Coherence, they say the 27B Atlas leads Vector by 0.5 points at most. These are their own results. Other people have published 1 quantized version.

Title: AgentDecisionBench L2 score
Common: 89.6%
Shift: 88.1%

• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own hardware, so decisions need no outside service.
• Not for: anything the card's own intended use and limitations section rules out, so read that section first.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4.1B
• Weights: about 8.2 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 10,582
• Likes: 44
• License: apache-2.0
• Based on: XHToken/Spark-X2.6-4B
• Quantized versions: 1
• Tags: transformers, safetensors, spark2_5, text-generation, decision-making, agents, agent-safety, causal-inference, consequence-prediction, on-device, laplace, juneo

Links

• Model: https://huggingface.co/JuneoAI/Laplace-1-Vector
• Files: https://huggingface.co/JuneoAI/Laplace-1-Vector/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download JuneoAI/Laplace-1-Vector

Then follow the usage in its model card. Nothing in this kit runs by itself.