Rizzo Flow 4B
A local, Jev-compatible decision model from Rizzo AI Academy. XHToken's Spark-X2.5-4B with a merged typed-decisions LoRA, run on llama.cpp; yes/no, choice and score questions share one prefill of the state and are read from the answer-letter logits. No text is generated.
On the test split of LocalLLaMA/typed-decisions the Q8_0 build scores 64.8% with a calibration error of 0.112, against 57.4% (0.349) for the original Spark-X2.5-4B on the same RTX 5060 Ti; Jev's published figure there is 72.7%. The authors say the probabilities are not calibrated for a new domain and ship temperature scaling for your own data. The rank-16 LoRA was trained for one epoch on 28,321 public questions, none labelled by Jev. The repository holds the merged model as BF16 safetensors and as Q8_0, Q4_K_M and BF16 GGUF, plus the adapter. The server also offers /v1/decisions, with numeric answers and abstention; choice takes up to 26 options, and a question 8,192 tokens by default (1M native). The runtime has been public since 21 September on the original weights; this fine-tune followed on the 25th. Sibling: rizzo-flow-1.7b (54.4%). Some training configs are CC-BY-SA or CC-BY-4.0; check them before redistributing derivatives.
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 calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4B |
| Base model | xhtoken/spark-x2.5-4b |
| Maker | Rizzo AI Academy |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 64.8% |
| Reported latency | 195 ms p50 per five-question case on an RTX 5060 Ti (Q8_0); about 50 ms for one short decision |
Get the weights
pip install systemonemodels
systemone pull rizzo-ai-academy/rizzo-flow
The files are served from the maker's Hugging Face repository, rizzoaiacademy/rizzo-flow, and verified against the checksums recorded here.