Standard One
Standard Thinking's Apache-2.0 3B and 8B decision models on Ministral 3 that score bounded answers via /v1/systemone, with vision.
Overview
Standard One is a family of decision models from Standard Thinking, released on Hugging Face under Apache-2.0 at 8B and 3B sizes on Ministral 3, with a retained Pixtral vision tower that accepts image data URLs. Given a state and a bounded rubric, the model scores the supplied labels in one forward pass without decoding text and returns probabilities through POST /v1/systemone: choice selects among labelled options, noul is yes/no and score uses an ordinal scale. Merged BF16 weights, LoRA adapters with merge recipes, GGUF builds and server code (a jev-adapter in front of SGLang) are provided; v2.2 is the current version.
Probabilities are temperature-scaled, and the README reports that in an offline comparison with its untuned base, the 8B improved on four of six suites and declined on the public easy and hard tiers, that 8B beats 3B on public standard and hard tiers while 3B is faster, and that its 10 evaluation suites exclude JevBench items when choosing wording or temperatures. Training data covers nine languages, with performance varying by language, and no image-input decision benchmark is reported.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which labelled option fits this scenario? | — |
| Yes / no | Yes/no decision about the scenario | — |
| Score | Ordinal score questions | — |
How it works
- Serve the merged model with SGLang and the jev-adapter
- Send state (text or image) and typed questions to /v1/systemone
- Score all labels in one forward pass
- Apply thresholds to temperature-scaled probabilities
Strengths
- Open weights with adapters and GGUF
- Jev-compatible endpoint, aliasable as jev-latest
- Multilingual training and vision tower
Caveats
- Some declines versus base on public tiers
- No image-input benchmark reported
- Measurement paths differ across versions
Getting started
Serve the model with SGLang, then run jev-adapter --engine-url http://127.0.0.1:30000 --model standard-one-8b --alias jev-latest.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- SGLang + jev-adapter server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Standard Thinking's Apache-2.0 3B and 8B decision models on Ministral 3, with a retained Pixtral vision tower: text or images in, option probabilities out through `/v1/systemone`, with merged weights, adapters, GGUF builds, and server code
AnotiaWang/awesome-decision-models
Frequently asked questions
Is Standard One compatible with Jev clients?
Yes, its adapter serves /v1/systemone and can alias the model as jev-latest.
Which sizes exist?
8B and 3B, as merged weights and LoRA adapters.
Does it handle images?
It accepts image data URLs via the Pixtral tower, but no image benchmark is reported.
Related projects
- Lichen★ 58
Drop-in local replacement for Jev that serves /v1/systemone from open weights, including questions about images.
ServicebetaPython Local /v1/systemone server in Japanese docs with LFM, Sarashina and fine-tuned ModernBERT-Ja backends, including image input.
ServicebetaPythonHugging Face card for NeoHorse-Jev-4B: Apache-2.0 4B decision model for text or image-plus-text, served at /v1/systemone.
ModelbetaPythonOpen 27B multimodal decision model built largely by autonomous agents, serving choice, noul and score on a TypeSafe-compatible API.
ModelresearchPythonApache-2.0 Gemma 4 12B fine-tune for local typed decisions, chat and image input, served via /v1/systemone and chat completions.
ServiceresearchStatic GGUF quantizations by mradermacher of SargeDev's JEV-27B-VL-Uncensored vision decision model.
Model