PKU-KCL-Meteor-v1.5-4B
Decision-Making AI Model Based on Qwen3.5
2 min readupdated
PKU-KCL-Meteor-v1.5-4B is a specialized model from the Peking University Computational Knowledge Laboratory (KCL), built on the Qwen3.5-4B architecture (4.54 billion parameters). It is designed for structured state analysis and answering typed questions in a single pass, outputting probabilities for each option without generating text. The model focuses on choice tasks, yes/no evaluations, and level ranking. It is useful for developers and researchers building decision-making systems and automated request processing without unnecessary generation.
Key features
- Question types: Supports multiple-choice options (up to 255 choices), yes/no probabilities (noul), and ratings on a scale from 2 to 10 levels (score).
- Single pass: All options are listed in the prompt under single-token codes, and the readout module evaluates them at the final position without text generation.
- Truncation protection: Requests exceeding the maximum context length of 262,144 tokens are rejected rather than truncated.
- Hardware requirements: Runs text data on a single NVIDIA GPU in BF16 format with an FP32 reader.
What you can do
- Option analysis: Integrate the model into decision-making pipelines to select the optimal action from multiple alternatives with confidence scores.
- Safety evaluation: Use binary probabilities (noul) for rapid verification of whether actions comply with safety rules.
- Level ranking: Apply a 2 to 10 level scale for gradient-based evaluation of states or objects.
- Server deployment: Run the model as part of the Decision Index kit or set up a server to process requests at /v1/systemone.
Example
from meteor import Meteor
model = Meteor()
model.decide(
"The traffic light turned red as the car approached.",
{"action": {"type": "choice", "criteria": {"stop": "Stop", "go": "Keep going"}},
"safe": {"type": "noul"}},
)
Read more
DetailsBenchmarks: Decision Index 0.3 score of 69.43 (top results in tools: 88.4, language: 71.6).
Training and details: BF16 weights with an FP32 reader, file integrity check via SHA-256.
Base model: Qwen/Qwen3.5-4B.
License: CC BY-NC 4.0 for weights, Apache License 2.0 for code.










Comments