# ZooWork: instinct > A 4B decision model from ZooWork: a merged LoRA fine-tune of Qwen3.5-4B that reads a state and returns a probability per candidate for choice, yes/no (noul) and score questions in one forward pass, without generating text. Open weights, also on ZooWork's hosted API. - Page: https://systemonemodels.ai/zoowork/instinct - API: https://api.systemonemodels.ai/v1/models/zoowork/instinct - Download: `pip install systemonemodels && systemone pull zoowork/instinct` ## Facts | | | |---|---| | Maker | ZooWork (https://systemonemodels.ai/zoowork) | | Decides | choice, score, noul | | Architecture | instinct | | Base model | qwen/qwen3.5-4b | | Parameters | 4.0B | | Context | 8K tokens | | Licence | apache-2.0 | | Availability | Open weights + hosted API | | Released | 2026-09-29 | | Latest version | 2026.09 | ## Hosted API - Provider: ZooWork (ZooData) (https://instinct.zoowork.ai/) - Docs: https://instinct.zoowork.ai/docs/ - Input: $0.010/MTok · Output: Free ## Reported evaluation Suite: JevBench public set (231 items), maker's run with the official JevBench client; items also used for model selection. Numbers are the publisher's own. - Decision accuracy: 85.7% - Median latency: 62 ms ## Model card # Instinct Tuned 4B A 4B decision model from ZooWork: a merged LoRA fine-tune of Qwen3.5-4B that reads a state and returns a probability per candidate for choice, yes/no (noul) and score questions in one forward pass, without generating text. Open weights, also on ZooWork's hosted API. A rank-32 LoRA merged into Qwen3.5-4B, trained on about 9.4k items (data not released). The readout is over the candidate label-token logits with a temperature per question type; it needs ZooWork's reference runtime rather than generate(), transformers 5.16.1 and a CUDA BF16 GPU. Text only, 8,192-token limit, mostly English, sensitive to option order; the maker advises refitting the temperature on your own data (hard-split ECE 0.04 to 0.11 depending on option order). JevBench is a third-party benchmark; the maker's run on its public set gives 85.71%, and the maker says those items were also used for model selection, so the score is not held out. Hosted siblings with no weights of their own: instinct (frozen Qwen3.8-27B with a logit readout, training-free; maker's run 87.45%) and instinct-dual-4b (frozen Qwen3.5-4B, two option orders averaged; 82.25%). Paid API through ZooData at $0.01 per million input tokens for the 4B models and $0.03 for instinct, output free; a free preview endpoint exists. Previously skipped on 30 Sep, when the only source was a third-party issue; the maker's own repo, card and site now exist. ## 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 ## At a glance | | | |---|---| | Parameters | 4B | | Base model | `Qwen/Qwen3.5-4B` | | Maker | ZooWork | | Released | 2026-09-29 | | License | apache-2.0 | | Reported accuracy | 85.7% | | Reported latency | 62 ms p50 direct upstream on one H200 (four replicas, warmed, serial); p95 about 110 ms, estimated by the maker | ## Hosted API Served by **ZooWork (ZooData)** — $0.01/MTok input, $0/MTok output. [Get access](https://instinct.zoowork.ai/) · [API docs](https://instinct.zoowork.ai/docs/). ## Get the weights ```bash pip install systemonemodels systemone pull zoowork/instinct ``` The files are served from the maker's Hugging Face repository, [`srpone/instinct-tuned-4b`](https://huggingface.co/srpone/instinct-tuned-4b), and verified against the checksums recorded here. ## Read more - [Model card](https://huggingface.co/srpone/instinct-tuned-4b) - [Source and runtime](https://github.com/SerendipityOneInc/instinct) - [Instinct site and hosted API](https://instinct.zoowork.ai/) --- *This page was opened by System One for ZooWork, who can claim the organisation and take it over at any time.* --- From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models