Jev directory
CommunityJev Models & Reimplementations
Standard One 3B LoRA
The v2.2 PEFT LoRA adapter for Standard Thinking's Standard One 3B decision model, shipped with a merge recipe for Ministral-3-3B-Instruct; serving requires the merged checkpoint and the shared server code in the StandardOne-8B repository.
- Reading time
- 2 min
- Published by
- Community
- Author
- Standard Thinking
- Last verified
- Oct 2026
- Platforms
- API
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 3B LoRA adapter (PEFT) with a merge recipe; serving needs the merged checkpoint and the server code in StandardOne-8B.
- Merges onto base mistralai/Ministral-3-3B-Instruct-2512-BF16 via PeftModel.merge_and_unload().
- Same endpoint as the merged model: choice, noul and score questions return probabilities, usage.output_tokens is 0.
- v2.2 continues training from v2.1; labels for more than 26 options run A-Z then AA, AB, ...
- Base, adapter, merged weights and shared server code are all Apache-2.0.
Quickstart
python
import torch
from transformers import Mistral3ForConditionalGeneration
from peft import PeftModel
base = Mistral3ForConditionalGeneration.from_pretrained(
"mistralai/Ministral-3-3B-Instruct-2512-BF16",
revision="b6d637bef2393152b3da2b2fde72eecdee30557e",
torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, ".").merge_and_unload()
model.save_pretrained("./StandardOne-3B-merged", safe_serialization=True)Watch out
Card reviewed 2026-10-17; this repository holds the adapter and merge recipe rather than runnable weights, and the published accuracy tables are the publisher's own served measurements of the merged model through SGLang 0.5.20 and jev-adapter. The wording temperatures come from a v2.1 measurement and were not refit for v2.2.
Questions about this entry
Short answers drawn from the linked project or page.
The v2.2 LoRA adapter for Standard Thinking's Standard One 3B decision model, distributed as PEFT weights with a merge recipe. It targets base mistralai/Ministral-3-3B-Instruct-2512-BF16 and needs the merged checkpoint plus the shared server code to serve typed decisions.
Load the base with Mistral3ForConditionalGeneration.from_pretrained at revision b6d637bef2393152b3da2b2fde72eecdee30557e, apply PeftModel.from_pretrained(base, ".").merge_and_unload(), save the merged directory, then copy tokenizer, chat template, processor and generation files from the base snapshot.
No. Serving requires the merged checkpoint and the server code in the StandardOne-8B repository; the card's quick start runs SGLang 0.5.20 for the engine and jev-adapter in a separate virtual environment.
choice (labeled options), noul (yes/no) and score (ordinal) questions return probabilities for the caller's labels in one forward pass without decoding answer text; usage.output_tokens is 0.
The serving example uses --prompt-wording served --label-scheme upper --default-temperature 1.65, the wording used for the v2.2 measurements; the card's native and served wording temperatures come from a v2.1 measurement and were not refit for v2.2.
Ministral-3-3B-Instruct-2512-BF16 at revision b6d637bef2393152b3da2b2fde72eecdee30557e. The base, adapter, merged weights and shared server code are all Apache-2.0.
It continues training from v2.1 with additional decision data, and questions with more than 26 options now use the labels A-Z followed by AA, AB and so on; both versions were measured the same way and the numbers are dated 3 October 2026.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Both Jev 1.13 and 1.14 score 29 of 30 on 30 Turkish buyer and seller messages
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 NewsMore Jev resources like this
FRIDA-Decisions answers Russian text with structured decisions in a single encoder pass — choice, ordinal score, yes/no or ranking — and scores 0.893 on the 735-item razvilka benchmark, effectively level with TypeSafe Jev's 0.897 (McNemar p = 0.84).
Jev Models & Reimplementations#community#open-models#russian
Community
Standard Thinking's merged BF16 3B checkpoint of Standard One: it scores choice, noul and score questions and returns probabilities through POST /v1/systemone with no generated text. The v2.2 card reports 100.00% JevBench public easy and 79.97% on many-option questions.
Jev Models & Reimplementations#community#open-models#multilingual
Community
The Ollama build of iapp's OpenThai-SystemOne: a 0.8B Thai and English decision model whose v0.3 weights are fine-tuned for Ollama's /v1/systemone letter readout, reporting a 74.3 macro across 13 public subsets for Q8_0.
Jev Models & Reimplementations#community#open-models#local
Community
Standard Thinking's merged BF16 8B checkpoint of Standard One, in the repository that also carries the shared server code: bounded choice, noul and score questions get probabilities through POST /v1/systemone, with no generated answer text.
Jev Models & Reimplementations#community#open-models#multilingual
Community
Standard Thinking's 8B LoRA adapter for the Standard One decision model: merge it with the supplied recipe and serve it, and it scores choice, noul and score labels over a scenario with calibrated probabilities via POST /v1/systemone, never generating text.
Jev Models & Reimplementations#community#open-models#local
Community
Nox-4B, the 4B model in vLLM Semantic Router's Decision 1.0 family: give it a state, questions and possible answers, and it returns typed decisions and probabilities with labels defined at runtime. The card reports 73.09% weighted accuracy across 3,766 decisions.
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.