LightDec_V2
A 160M-parameter encoder decision model from Falcons.ai on the Ettin-150M backbone. It reads a state of up to 2,048 tokens and answers choice, yes/no (noul) and ordinal score questions in one pass per question, with a temperature-calibrated probability per option.
FalconDec reads the question, every option and the state in one encoder pass; a set-transformer head that ignores option order compares the options, and a temperature per question type and option count calibrates the result, with a defer flag. English only; it needs the falcondec_modeling.py loader in the repo. The maker's numbers come from its own test split, whose tasks are drawn from the same public sources as its training data; on tasks it did not train on accuracy is lower (51.8% on the TEV1 tasks with no source overlap). The card says it is weak at arithmetic, table counting and maths or knowledge tasks, and that prompt-injection detection is not reliable enough to be a sole safety layer; the 100% long-context scores are synthetic. Siblings: LightDec (v1, 512 tokens), LightDec_Arthur (12M byte-level) and Athr_Agent_Sec (an agent-step monitor). The repo also holds Model Surgeon packaging files that are not needed to run the model and are not listed here. Falconsai/laya-v796 is a re-upload of Convai's Laya and is not listed.
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 | 160M |
| Base model | jhu-clsp/ettin-encoder-150m |
| Maker | Falcons.ai |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 78.4% |
| Reported latency | 10.05 ms p50 per call (one question) on an NVIDIA RTX PRO 6000 Blackwell, fp16; 48.4 ms on CPU with the int8 weights |
Get the weights
pip install systemonemodels
systemone pull falcons-ai/lightdec
The files are served from the maker's Hugging Face repository, Falconsai/LightDec_V2, and verified against the checksums recorded here.