AnyS1-4B
a fast System-1 model for decisions in agentic workflows
2 min readupdated
AnyS1-4B is built on Qwen/Qwen3.5-4B-Base using LoRA (a lightweight add-on on top of a base model) and a custom OptionInteractionHead. It evaluates all possible response options for an agent in a single pass and returns calibrated probabilities. A confidence threshold decides whether to answer locally using this model or escalate the task to a larger System-2 model. It helps AI agent developers cut latency and reduce the cost of calling heavy LLMs.
Key features
- Built on
Qwen/Qwen3.5-4B-Basewith a LoRA adapter (rank 16) and a dedicated option-selection head. - Delivers high benchmark accuracy (83.8% on Offline Dev and 93.5% on SOPBench E nodes).
- Achieves a median single-pass latency of 21.8 to 31.7 ms on a single H100.
- Cuts episode execution costs by 75–91% and reduces model runtime by 63–92% compared to a standard setup.
What you can do
- Spin up a local server to evaluate AI agent decisions using the official repository.
- Integrate the model into custom agentic workflows for fast option filtering and classification.
- Tune the confidence gate (confidence threshold) to flexibly route between the local model and System-2.
Example
Starting a local inference server:
git clone [https://github.com/middle-chunjie/AnyS1.git](https://github.com/middle-chunjie/AnyS1.git) && cd AnyS1 && pip install -e .
huggingface-cli download smSquirrel/AnyS1-4B --local-dir ckpt/AnyS1-4B
python -m anys1.model.serve_s1 --checkpoint ckpt/AnyS1-4BRead more
DetailsBase model: Qwen/Qwen3.5-4B-Base
Training: 25,986 states across 1,072 public skills
Accuracy: 83.8% Offline Dev, 93.5% SOPBench E nodes
Limitations: works only with enumerable decisions; free-form generation falls back to System-2










Comments