Super Decider, an open-weight DiffusionGemma model, takes #1 on the JevBench API board
Source · Model card: Super Decider
publicado (OK do Fabricio 26/10 BRT).
Benchmark Heaven’s JevBench API board now lists Super Decider from Singularity Compute at #1 of 48 hosted decision APIs, with a composite of 78.5. It was measured on 21 October under JevBench v1.7.0, which counts any decision over 10 seconds as wrong. ASI:Cloud, which hosts the model, announced the ranking on 26 October at 07:35 BRT.
| Super Decider | kettei-fat | Fastino GLiDE | Jev 1.13 | |
|---|---|---|---|---|
| Composite | 78.5 (#1) | 77.7 (#2) | 76.4 (#3) | 71.4 (#14) |
| Intelligence | 78.9 | 79.8 | 76.1 | 63.6 |
| Calibration | 90.8 | 92.8 | 92.8 | 90.6 |
| p50 latency | 0.25 s | 0.20 s | 0.36 s | 0.24 s |
| $ per 1,000 decisions | 0.016 | 0.026 (est.) | 0.023 | 0.032 |
Super Decider wins on the composite because it is cheaper, not because it is the most accurate: kettei-fat has the higher Intelligence and the top Capability Score (86.3 vs 84.9). With thinking forced on, Super Decider reaches Intelligence 80.0 but drops to #8 (74.5) because its median latency rises to 0.89 s.
It is the only top-six row that the board tags as open weights. The Hugging Face repo (Apache-2.0, 13 Oct) holds an 18 GB NVFP4 fine-tune of google/diffusiongemma-26B-A4B-it, and the runtime serves POST /v1/systemone in the Jev request shape. On ASI:Cloud it costs $0.03 per million input tokens, with output free.
Caveats: this is one board, and API rows are measured on endpoints that see the sealed inputs. The model has no row on the Decision Index or on Artificial Analysis yet. Training details (“permissively licensed and synthetic decision data”, no test items) come from the vendor.
More: Super Decider.