typical-medium
Oz Labs' open decision model. Give it a state, a question and options set at call time; it returns a probability per option plus an explicit abstain, in one forward pass with nothing generated. Weights are a LoRA adapter with Choice, Score and Noul heads over Qwen3.5-4B.
The state is KV-cached, so further questions about the same state are cheap (about 2.7 ms each for typical-small, per the maker). The weights were replaced in place on 23 September (v2, Qwen3.5-4B-Base); the card front matter, MANIFEST.json and JevBench summary still describe v1, and an older inference package will not load v2. The v2 swap cost 5.5 points on CLINC-150 and 5.3 on HWU64, which the maker reports. Maker's own numbers for v2: .950 on 605 held-out long states, .858 held-out yes/no, .586 held-out Score; on the public subset of the third-party JevBench, its own run gives .861 standard and .495 hard, so the hard tier is weak. Calibration is the maker's claim, not independently checked. The maker flags that one training set (metaeval/ambient) has no declared licence. best.pt is a PyTorch pickle holding the adapter and heads; the Qwen base downloads separately. Siblings: typical-small (1.7B on Qwen3-1.7B-Base), typical-small-preview and typical-large-preview (14B, a research preview that misses the maker's own release bar).
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-Base |
| Maker | Oz Labs |
| Released | 2026-09-24 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull oz-labs/typical
The files are served from the maker's Hugging Face repository, OzLabs/typical-medium, and verified against the checksums recorded here.
Read more
This page was opened by System One for Oz Labs, who can claim the organisation and take it over at any time.