SynACK Decide 26B-A4B
A Gemma-4-26B-A4B decision model by Benjamin Babeshkin (SynACK): a JSON state and typed choice or yes/no questions go in, and one forward pass returns a probability for every option, with no generated text, over the /v1/systemone wire format.
About 4B parameters active per token; 51.6 GB in bf16, with GGUF (Q8_0, Q4_K_M) and 8-bit MLX builds from the maker. Up to 255 options; score was not trained. Calibrated in-distribution only and can be overconfident on unfamiliar compositions; mostly English. The card says it was built with Claude in Claude Code. Training data includes non-commercial sets (ANLI, Financial PhraseBank, RACE) and labels distilled from Kimi-K3; the maker says to review these before commercial use. The maker's own run of the third-party Decision Index 0.2.1 gives 59.5 (not verified by the board). SynACK is an independent project, not affiliated with Synack, Inc.
What it decides
- choice — picks one option from a set
- 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 | 26B |
| Base model | google/gemma-4-26B-A4B-it |
| Maker | Benjamin Babeshkin (SynACK) |
| Released | 2026-10-04 |
| License | apache-2.0 |
| Reported accuracy | 76.8% |
| Reported latency | 132 ms median / 360 ms p95 per Decision Index request, bf16, one RTX PRO 6000 |
Get the weights
pip install systemonemodels
systemone pull benjamin-babeshkin/synack-decide
The files are served from the maker's Hugging Face repository, SkyPanther/synack-decide-26b-a4b, and verified against the checksums recorded here.
Read more
This page was opened by System One for Benjamin Babeshkin (SynACK), who can claim the organisation and take it over at any time.