basal-1.5
Polish and English typed-decision models fine-tuned from SpeakLeash's Bielik. A state goes in; choice, yes/no (noul) and score questions come back as a calibrated probability per option, read from the option-letter logits in one pass per option order, with no generated text.
Polish first, English second. The served answer averages two passes, one per option order, then applies a per-type temperature; the confidence thresholds are validated only on some prompt types. The bundled engine (rkinas/basal) serves /v1/systemone and adds experimental multi, act and evidence-span types; it runs on PyTorch, vLLM, SGLang, MLX, Ollama and llama.cpp. Up to 10 options per question. Siblings, all Apache-2.0: basal-1.5-max (11B, the maker's most accurate), basal-1.5-mini (1.5B) and basal-1.0 (26 September, with a technical report on Zenodo). The maker's own numbers: 0.931 on a sealed Polish test, 0.688 on ten public English tasks; Werdykt is the maker's own hidden benchmark, with only samples public.
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
- route — sends the input to one of several destinations
At a glance
| Parameters | 4.5B |
| Base model | speakleash/Bielik-4.5B-v3.0-Instruct |
| Maker | Remek Kinas |
| Released | 2026-10-05 |
| License | apache-2.0 |
| Reported accuracy | 72.1% |
| Reported latency | 33.8 ms median per decision with SGLang on one H100 |
Get the weights
pip install systemonemodels
systemone pull remek-kinas/basal
The files are served from the maker's Hugging Face repository, Remek/basal-1.5-4.5B, and verified against the checksums recorded here.