Jev directory
CommunityJev Models & Reimplementations
Super-Decider
Singularity Compute's fine-tune of Google's diffusiongemma-26B-A4B-it, a discrete-diffusion mixture-of-experts, served from an 18 GB NVFP4 checkpoint that answers choice, noul and score questions with calibrated distributions on one 48 GB GPU.
- Reading time
- 3 min
- Published by
- Community
- Author
- Singularity Compute
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- API
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Fine-tuned DiffusionGemma 26B-A4B (26B total, 4B active) merged into NVIDIA's 18 GB NVFP4 checkpoint with 4-bit experts.
- Returns choice probabilities, P(yes) for noul and a score expectation over an ordered rubric, one answer per question per read.
- One denoising step fills all answer slots, so several questions cost little more than one; contexts run up to 131,072 tokens.
- Serves POST /v1/systemone and OpenAI-compatible chat completions on one 48 GB GPU; L40S runs NVFP4 via Marlin W4A16.
- Adaptive thinking rereads an answer after a short thought when unsure; Apache-2.0, and no benchmark test items were used for training.
Watch out
Card reviewed 2026-10-17; the card publishes no accuracy or calibration benchmarks, so this entry rests on the architecture and serving description alone. The card itself warns that answers can change with meaning-preserving input changes, that the model is sometimes confidently wrong on multi-step date and money calculations, and that training data is English-centric with no image fine-tuning.
Questions about this entry
Short answers drawn from the linked project or page.
A decision model by Singularity Compute: given a state (text, JSON or images) and typed questions, it returns one calibrated answer per question in a single read, built on the discrete-diffusion mixture-of-experts google/diffusiongemma-26B-A4B-it.
Choice, returning a probability per option; noul, returning P(yes); and score, returning the expected level over an ordered rubric together with the distribution.
All answer slots of a request are filled in one denoising step, so several questions cost little more than one. Each question's prompt block is read with bidirectional attention through a small vLLM patch (DJEV_BIDIR_SERVE=1, DJEV_BIDIR=qsep), and adaptive thinking can be turned off, kept on under load or forced per question.
Use the Singularity-Compute/super-decider runtime: a pinned vLLM nightly with patches and a structured server that accepts the Jev/System One request format at POST /v1/systemone and OpenAI-compatible chat completions. One 48 GB GPU suffices, and on L40S the NVFP4 weights run through the Marlin W4A16 kernel.
It fine-tunes google/diffusiongemma-26B-A4B-it and is served from NVIDIA's 18 GB NVFP4 checkpoint with 4-bit experts. Contexts go up to 131,072 tokens, and generation_config.json sets a confidence threshold of 0.3 for thought canvases.
Answers can change when the input changes in ways that preserve meaning, such as JSON key order or option order; the model is sometimes confidently wrong on multi-step date and money calculations, which thinking on every question fixes only partly; and training data is English-centric, with images handled by the base vision path without image fine-tuning.
Apache-2.0, like the base model, per the card. It accepts the Jev / System One request format and mirrors the typed choice, noul and score contract, but it is an independent community model rather than TypeSafe Jev.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
A LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
Jev Models & Reimplementations#community#open-models#local
Community
From 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.