super-decider

What it does. Given a state (text, JSON or images) and typed questions, the model returns one calibrated answer per question in a single read: choice (a probability per option), noul (P(yes)) and score (expected level over an ordered rubric, with the distribution).

Model

• Base: google/diffusiongemma-26B-A4B-it (discrete-diffusion MoE, 26B total, 4B active), served from NVIDIA's NVFP4 checkpoint nvidia/diffusiongemma-26B-A4B-it-NVFP4.
• 18 GB NVFP4 checkpoint (experts 4-bit). No benchmark test items were used for training.
• generation_config.json sets confidence_threshold 0.3 for thought canvases.
• Served with contexts up to 131,072 tokens.

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. All answer slots of a request are filled in one denoising step, so several questions cost little more than one.
• Bidirectional attention within each question. The prompt block of every question is read in both directions, not only left to right. The model was trained this way, and the runtime applies the same attention through a small vLLM patch (DJEV_BIDIR_SERVE=1, DJEV_BIDIR=qsep).
• Adaptive thinking. When the model is unsure of an answer, it writes a short thought and reads the answer again. A request can turn this off, keep it on under load, or think on every question with "thinking" (see API docs).

Serving

Runtime: Singularity-Compute/super-decider (pinned vLLM nightly, vLLM patches and a structured server). It accepts the Jev / System One request format (POST /v1/systemone) and OpenAI-compatible chat completions. One 48 GB GPU is enough; on L40S the NVFP4 weights run through the Marlin W4A16 kernel.

Limitations

• Answers can change when the input changes in ways that keep its meaning, such as JSON key order or option order.
• The model is sometimes confidently wrong on multi-step date and money calculations; thinking on every question fixes part of them.
• The training data is English-centric. Images go through the base model's vision path without image fine-tuning.

Licence

Apache-2.0, like the base model. See NOTICE.md.