Jeff v1.3 (jeff-base)
A 0.8B fine-tune of Qwen3.5-0.8B that returns a calibrated probability for each listed option in one forward pass (choice, yes/no, score), through a trained readout head. v1.3 is adapter-first, with 15 per-task LoRA adapters; for zero-shot use the maker points to v1.2.
A personal project by Mathias Strasser, started from Denis Yarats' AutoJev recipe; not affiliated with TypeSafe and unrelated to other models named Jeff. English text only. Without an adapter the v1.3 base is weak on long, unfamiliar option lists (7.4% on legal clauses, against 66.0% for v1.2), so the zero-shot v1.2 builds (jeff-legacy/Jeff-Qwen3.5-0.8B and 2B) stay published. The adapters differ in licence: sanctions and soc are CC BY-NC 4.0 (never packaged here), and aml and trading-desk carry data-source terms of their own; Jeff-Gemma4-E2B follows the Gemma 4 terms. The training data mixes sources, some under CC BY-SA. Each adapter's held-out accuracy is on its own card (for example guard 98.2%, triage 91.5%). Python and TypeScript clients, GGUF builds and a jeff-serve server that switches adapters per request.
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 | 0.8B |
| Base model | Qwen/Qwen3.5-0.8B |
| Maker | Mathias Strasser |
| Released | 2026-10-05 |
| License | apache-2.0 |
| Reported accuracy | 78.6% |
| Reported latency | 26.6 ms median per decision, base alone, on an RTX PRO 6000 via jeff-serve (675 requests); 31.4 ms with one adapter |
Get the weights
pip install systemonemodels
systemone pull mathias-strasser/jeff
The files are served from the maker's Hugging Face repository, mstrasser/jeff-base, and verified against the checksums recorded here.