arc-decide-0.6b-v0.1
A 0.6B decision model on Qwen3-0.6B-Base from the DocsGPT team. Give it a state and typed questions (yes/no, a choice of up to 16 options, a 3- or 4-level score) and it returns calibrated probabilities in one pass, for RAG and agent checks.
Typed linear heads are folded into the LM head, so stock vLLM can serve it with one token plus a bias and temperature per head from arcd_serving.json; a temperature is fitted per canonical question (retrieval relevance and grade, guardrails, routing, conversation and tool analytics). English, Spanish, Chinese and Russian. The numbers are Arc53's own on its held-out harness, and the analytics gold labels are agreement between two LLMs. Injection detection is weaker on novel attacks (12% false positives on NotInject at 0.5); the maker advises a threshold of at least 0.9 and not using it as the only security control. Questions far from the canonical ones are much less reliable. A CPU ONNX export is in the repo. The training data includes CC BY-SA sets.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- rank — orders a set of items
- route — sends the input to one of several destinations
At a glance
| Parameters | 0.6B |
| Base model | Qwen/Qwen3-0.6B-Base |
| Maker | Arc53 |
| Released | 2026-09-30 |
| License | mit |
Get the weights
pip install systemonemodels
systemone pull arc53/arc-decide
The files are served from the maker's Hugging Face repository, Arc53/arc-decide-0.6b-v0.1, and verified against the checksums recorded here.