Laplace-1-Atlas

Laplace-1-Atlas

Helps organizations decide who does a task, who sees information, what runs and what gets accepted, by scoring candidate options in one pass. A 27.4B model built on Qwen3.8-27B with open Apache-2.0 weights.

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-Atlas is the larger of two Laplace-1 models from JuneoAI, built on Qwen3.8-27B with 27.4B parameters. It is a decision model that predicts the effect an action would have inside an organization, then helps pick among candidates. You hand it a set of options and read a probability spread over them, with no sampling and no parsing.

Who it suits: It suits teams that let software agents take over work and want a check on each choice before it runs. The card frames four kinds of decision.

Industry | Who uses it | How they use it | Example
Sales operations | Teams that route leads | Allocate work by picking who should own a task | Grade a new lead and pick the people best placed to follow it up
Data governance | Security and compliance teams | Decide who may see which information | Limit a client's private details to the roles that need them
Operations | Teams that automate processes | Decide whether to run, adjust or ask first | Send a repair order straight away, but route a large contract to its owner
Delivery | Quality and review teams | Check results against a standard | Compare a deliverable with its acceptance criteria before sign-off

You can skip it if you want a chat or writing model, since the card builds it for decisions.

Where to use it:

• Organization servers and private cloud, which the card names for this larger model
• Python with the transformers library
• Windows, macOS and Linux with Python, to download the files
• Agent systems that need to score candidate actions before they run

Use case | Who | What they get back
Task allocation | Operations teams | A probability for each candidate owner
Access decisions | Security teams | A probability for each sharing option
Action choice | Automation teams | A spread over run, adjust or ask first

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• Disk and memory for about 54.7 GB of BF16 weights, or about 13.7 GB for a 4-bit copy, which is an estimate
• The transformers library

1. Install or update the Hugging Face download client.
2. Download this repository's files to your computer.
3. Open the Usage section of the model card for the loading steps, which are not in the text available here.

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

How to use it: You give the model the state of an organization, the chain of authority behind an agent and a set of candidates. You read the probabilities over those candidates from the next-token logits.

Tool | What it does
pip install -U huggingface_hub | Installs or updates Hugging Face's download client
hf download | Fetches a model repository's files to your computer

Example: You describe a new task and list three possible owners. The model scores each owner, and you read the spread of probabilities to see who it favors. The card gives no sample output here, so check the Usage section for the exact format.

For newcomers and experts: Think of it as a helper that picks among options you supply, such as who should handle a task or whether an action should run. A first small step is to read the Usage section, then fetch the files with the two commands above.

For readers who know the field, the card says it learns the interventional distribution of an organization's next state after a candidate action, given the chain of authority behind the agent. A decision is one forward pass over candidates. The numbers below are the authors' own, so check them on your data.

Describe the organization and chain of authority
List the candidate actions
Run one forward pass
Read probabilities from next-token logits
Choose, or send for review

Good to know:

• Claims: the card says first on 7 of 9 decision benchmarks, with ties counted, and 89.4 percent on TrustDecision.
• Sibling: Laplace-1-Vector has 4.1B parameters and targets on-premises and on-device use, where data never leaves.
• Weights: complete BF16 weights and tokenizer, with no adapter and no gated access.
• Sources: the card says the training data comes from a project it calls Decision Forge.

Title: Score on AgentDecisionBench L2
common: 95.6%
shift: 95.5%

• License: Apache 2.0.
• Cost to run: none named. You need enough hardware for the 54.7 GB of weights.
• Data it reaches: the organization state and candidate actions you pass in. The text available here names no network use.
• Not for: chat or writing, since the card builds it for decisions.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 10,692
• Likes: 51
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 5
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, decision-making, agents, agent-safety, causal-inference, consequence-prediction, laplace, juneo, text-generation

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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