Jev directory
CommunityJev Models & Reimplementations
StartLux-Decision v1.1 27B AutoThink
StartLux Labs' 27B AutoThink decision model reads a state (text, JSON or up to 8 images) and typed questions, answers in one pass and only writes a thought when the top probability falls below the 0.6 gate; self-reported Decision Index 0.3 public score 73.37.
- Reading time
- 3 min
- Published by
- Community
- Author
- StartLux Models
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- API
- Capabilities
- Self-hostedRuns locally
Highlights
- Self-reported Decision Index 0.3 public score 73.37 and DecisionPivot pair accuracy 86.9, against 72.83 and 75.4 for the 27B Instant sibling.
- Writes a thought of at most 4,096 tokens only when the top option probability is below the gate: 0.6 by default, 0.9 in the reported results.
- On DecisionPivot a 0.9 gate spends 304 thought tokens per question and reaches 86.9% pair accuracy against 48.0% with thinking off.
- vLLM BF16 median read-out is 49 ms at 256 prompt tokens and 698 ms at 8,192 on one NVIDIA H200; MLX 8-bit runs on an Apple M5 Pro.
- Stage 3 trains the thought with supervised reasoning examples and RLCD squared; weights are CC BY-NC 4.0 and the code is Apache-2.0.
Watch out
Card reviewed 2026-10-17; the Decision Index and DecisionPivot numbers are StartLux's own runs of its public kit and benchmark, the card states they reflect model selection rather than independent test results, and AutoThink rows use a gate of 0.9 rather than the 0.6 default. The weights are CC BY-NC 4.0, and MLX BF16 does not fit a 64 GB Apple machine.
Questions about this entry
Short answers drawn from the linked project or page.
A 27B decision model from StartLux Labs that answers typed questions and reasons only where it is unsure. It reads a probability for every option in one pass, and when the top probability is below the gate it writes a short thought, reads again and mixes the two read-outs.
Each question is read as its own prompt holding the whole state. If the top option probability reaches the gate (0.6 by default), that read-out is the answer; otherwise the same weights write up to 4,096 tokens of thought, read the letters again and answer 0.25 times the first read-out plus 0.75 times the second.
Choice questions with several options, noul yes/no questions and score questions on a scale of 2 to 10, over a state of text or JSON with up to 8 images. The card says the request and response format is compatible with the System One /v1/systemone API.
The quick start installs the package with vLLM extras and serves through startlux-decide serve --launch-vllm. Latency was measured on one NVIDIA H200; images and prompts over 65,536 tokens need vLLM 0.31, and several GPUs can each run a server behind the same decision server.
Decision Index 0.3 public 73.37 and DecisionPivot pair accuracy 86.9 with a 0.9 gate, against 48.0% with thinking off, at 304 thought tokens per question. The card says the scores are its own runs of the public kit compared across recipes and seeds, so they are not independent test results.
StartLux Labs built it, with code at StartLuxLabs/StartLux-Decision and a technical report at arXiv 2610.17309. The weights are CC BY-NC 4.0, so research and other non-commercial use is allowed and commercial use needs a separate licence.
The Instant sibling always answers from one read-out, while AutoThink writes a thought only when confidence is below the gate. AutoThink reports higher pair accuracy (86.9 against 75.4) and a slightly higher public index (73.37 against 72.83), at the cost of extra latency when it thinks.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Standard One SH (StandardThinking)
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
prithivMLmods' GGUF builds of autotrust/JEV-35B, the Qwen3.5-35B-A3B MoE decision model that returns calibrated choice, noul and score answers over text and images; files run from BF16 to Q3_K_L for llama.cpp.
Jev Models & Reimplementations#community#open-models#local
Community
A 4-bit W4A8 AWQ plus GPTQ build of AutoTrust's JEV-27B-VL from AI Outfitters that keeps the rank-32 LoRA decision head and a bf16 vision tower, served by vLLM on a single 24 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
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.
Jev Models & Reimplementations#community#open-models#local
Community
GGUF quantizations of StartLux Labs' 2B Instant decision model for llama.cpp: Q8_0 (2.01 GB) and Q4_K_M (1.27 GB), plus a 668 MB F16 vision encoder, with Q8_0 scoring 23.7 DecisionPivot against 23.5 for BF16 on vLLM.
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.