singularity-compute/super-decider
Organization repository on GitHub
super-decider: typed decisions (Jev-compatible) on fine-tuned DiffusionGemma 26B-A4B
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
Super Decider
Typed decisions on a fine-tuned DiffusionGemma 26B-A4B: super-decider answers choice, noul (yes/no) and score questions with calibrated distributions in one call, speaks the Jev/TypeSafe System One request format and OpenAI chat completions, and runs on one NVIDIA L40S. NVFP4 weights are 18 GB.
- Stars
- 5
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- singularity-compute
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Fine-tune of Google's DiffusionGemma 26B-A4B — a discrete-diffusion MoE with 4B active parameters — merged into an NVFP4 checkpoint.
- Reads every answer slot in one denoising step; reference hardware is a single NVIDIA L40S.
- Speaks the Jev/TypeSafe System One format at POST /v1/systemone plus OpenAI-compatible chat completions.
- Accepts text, JSON and up to 8 images per request; returns choice, noul and score answers with calibrated distributions.
- Runs on stock vLLM (pinned nightly) with small patches: bidirectional question prompts, faster Ada attention, a faster thinking path.
Watch out
README reviewed 2026-10-17; the digest truncates the How-it-works section mid-bullet, so the training-data composition, evaluation results and latency figures are not captured. The vLLM build is pinned to a nightly, and the NVFP4 weights (18 GB) live on Hugging Face rather than in the repository.
Questions about this entry
Short answers drawn from the linked project or page.
A fine-tune of Google's DiffusionGemma 26B-A4B that answers typed questions about a state — choice, noul (yes/no) and score — each with a calibrated probability distribution, in one call and one denoising step.
They are the Jev/TypeSafe System One question shapes: choice picks among options, noul is a yes/no judgement and score rates on a scale. super-decider returns a calibrated distribution for every slot rather than prose.
Through the Jev/TypeSafe System One request format at POST /v1/systemone, or through OpenAI-compatible chat completions. Requests can carry text, JSON and up to 8 images; docs/API.md is the stated reference.
One NVIDIA L40S per the README, running stock vLLM on a pinned nightly with small patches. The model is a 26B-A4B MoE merged into an NVFP4 checkpoint, so it needs an Ada-class GPU with enough memory for the 18 GB weights.
A discrete-diffusion mixture-of-experts model from Google with 26B total and 4B active parameters. super-decider fine-tunes it on permissively licensed and synthetic decision data, then merges it into NVIDIA's NVFP4 checkpoint.
NVFP4 weights and a model card are on Hugging Face under Singularity-Compute/super-decider, 18 GB. The repository code is Apache-2.0 with a NOTICE.md; the digest does not spell out separate weight licence terms.
The Singularity-Compute organisation. Metadata: Python, Apache-2.0, 5 stars, created 2026-10-13 and pushed 2026-10-15. The digest shows no benchmark table, so its calibration claims are not independently verified here.
More Jev resources like this
A small, fast decision model — a system one model — for agentic workflows from Strands Labs: it picks between options or rates on a scale faster than an LLM and returns a calibrated confidence with every decision.
Jev Models & Reimplementations#community#github#python
Community
strands-labs/strands-deciderPython
strands-labs/strands-deciderPythonAn open System One decision model from Interfaze-ai: give lev a state — text, a ticket, an email, or JSON — plus yes/no, choice, or score questions and it answers all of them in one forward pass from logits, returning calibrated probabilities and zero output tokens.
Jev Models & Reimplementations#community#github#python
Community
abhinavexists/levPython
abhinavexists/levPythonA Chinese-language decision model that returns a probability for each fixed candidate option in one forward pass, with no generated text or chain of thought. It is a LoRA fine-tune of Qwen3.5-0.8B on 13,048,000 CJ-Dataset rows, benchmarked on CJ-Bench against Chinese-Jev.
Jev Models & Reimplementations#community#github#python
Community
yanqiangmiffy/chinese-decisionPython
yanqiangmiffy/chinese-decisionPythonA training recipe for JEV-style System 1 decision models on Qwen bases: LoRA plus a 24-slot decision head distilled from Jev v1.13's output distribution, serving calibrated per-option probabilities through a TypeSafe-compatible endpoint.
Jev Models & Reimplementations#community#github#training
Community
shibing624/train-jevPython
shibing624/train-jevPythonThree open-weights 4B decision models for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B) that choose between options and return a probability for each. The server implements the Jev API, POST /v1/systemone, so existing Jev clients work by changing the endpoint.
Jev Models & Reimplementations#community#github#python
Community
mertkayacs/jevaltPython
mertkayacs/jevaltPythonAn open decision model built on frozen Qwen3.8-27B that processes a shared application state with isolated question branches and returns probabilities with zero generated answer tokens. It reports 87.45% on JevBench, exceeding Open-Jev-27B-v1.1 by 2.16 points, and needs 64 GiB of free GPU memory.
Jev Models & Reimplementations#community#github#python
Community
Camellia86/Canopy-JevPython
Camellia86/Canopy-JevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.