Decision 4B v1.5 (FlyMyJev-4B preview)
A LoRA on Qwen3.5-4B that reads option-letter logits in one pass and returns a probability for each declared option of a yes/no, choice or score question, through TypeSafe's /v1/systemone wire format. Adapter and code Apache-2.0; base downloaded separately. Preview.
A rank-16 adapter (57.5 MB); the base downloads separately at a pinned revision. The yes/no temperature (0.40) was chosen so that answers commit, not for calibration (ECE 0.112 on JevBench's public hard tier). Training data is not published; the card says only "our own decision data and licence-reviewed public datasets", and that no JevBench item and no Jev output was used. Needs a CUDA GPU and a C compiler at run time. v1.1 to v1.4 are earlier separate repos (v1.3 is the v1.2 weights with a new yes/no temperature). FlyMy.AI's 2B and Fast previews are evaluation-only and are not listed. Credits on the card: the prompt format is SemIf by TheoLeeCJ (MIT) and the recipe and CUDA-graph design come from JevK5 by allebee. JevBench is a third-party benchmark; the numbers are the maker's own package check. Not affiliated with TypeSafe.
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 | FlyMy.AI |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 88.3% |
| Reported latency | 18.9 ms p50 / 22.0 ms p95 per short decision on one RTX 4090 with CUDA graphs, in-process; 54 ms eager |
Get the weights
pip install systemonemodels
systemone pull flymy-ai/decision-4b
The files are served from the maker's Hugging Face repository, flymy-ai/decision-4b-v1.5, and verified against the checksums recorded here.