Super Decider model card
Overview​
| Property | Value |
|---|---|
| Developer | Singularity Compute |
| API model string | super-decider |
| Base model | google/diffusiongemma-26B-A4B-it, a discrete-diffusion mixture-of-experts model |
| Parameters | 26B total, 4B active |
| Weights | NVFP4 (experts in 4-bit), an 18 GB checkpoint |
| Context window | 131,072 tokens |
| Input | Text, JSON, up to 8 images |
| Output | Typed answers: choice, score, noul, span |
| Languages | English |
| License | Apache-2.0 |
| Weights download | Hugging Face |
| Serving runtime | GitHub |
Pricing on ASI:Cloud​
| Input / 1M tokens | Output / 1M tokens |
|---|---|
| $0.03 | Free |
See the model catalogue to compare with other models.
How it works​
- Fine-tuned for decisions. DiffusionGemma 26B-A4B is fine-tuned on permissively licensed and synthetic decision data, then merged into NVIDIA's NVFP4 checkpoint.
- One-step reads. The model fills all answer slots of a request in one denoising step, so several questions cost little more than one.
- Reads each question both ways. The prompt block of each question is read in both directions, not only left to right. The model was trained this way, and the serving runtime applies the same attention.
- Adaptive thinking. When the model is unsure of an answer, it can write a short thought and read the answer again. See Thinking.
No benchmark test items were used for training.
Reported results​
Singularity Compute reports these results for Super Decider, which costs $0.03 per 1M input tokens on ASI:Cloud, with output free:
- 85.6 on NVIDIA's When2Call, a benchmark of when an agent should call a tool.
- 94.5% accuracy checking documents of up to 16k tokens.
- +2 points on DecisionBench with thinking turned on.
- Ahead of TypeSafe's Jev on When2Call and PhishNChips.
- Ahead of Cloudflare's Clef on When2Call and PhishNChips, and 1.6 points behind it on BFCL tool selection.
- About 0.3 s per request over the public internet.
For comparison, the larger chat models that Singularity Compute prices against on its product page (Gemma 4 31B, gpt-oss-120b, MiniMax M3 and Qwen3.8 27B) cost $0.14 to $0.45 per 1M input tokens on ASI:Cloud, plus $0.40 to $3.20 per 1M output tokens. See the model catalogue.
These are the developer's figures. Measure the model on your own data before you rely on it.
Known limitations​
- Sensitive to input order. Answers can change when the input changes in ways that keep its meaning, such as the order of JSON keys or of a question's options. Build requests the same way every time.
- Multi-step arithmetic. The model is sometimes confidently wrong on multi-step date and money calculations. Thinking on every question fixes part of these; doing the arithmetic in code avoids them.
- English-centric. The training data is mostly English.
- Images. The model wasn't fine-tuned on images. It reads them through its base model's vision path, and image requests only think with
"thinking": "always".