Jev directory
CommunityJev Models & Reimplementations
StartLux-Decision v1 35B-A3B
StartLux's 35B-A3B mixture-of-experts decision model, answering typed questions through the TypeSafe /v1/systemone format with a probability per option; text, JSON or images up to 262,144 tokens, and a card-reported 210/231 on public JevBench.
- Reading time
- 3 min
- Published by
- Community
- Author
- StartLux Models
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Mixture of experts with about 35B parameters and about 3B active per token; the download is 72 GB and bf16 serving takes about 65 GiB of GPU memory.
- Answers every question with a probability per option in one forward pass over text, JSON or images up to 262,144 tokens.
- Speaks the TypeSafe /v1/systemone request and response format, so clients written for Jev work unchanged, per the card.
- Card reports Decision Index 0.2.1 61.55, Decision Index 0.2 57.24, JevBench public 210/231 and Intern-Decision 92.29 average over seven suites.
- On one H200 in bf16: 52.5 ms for a request with three questions answered in one forward pass, 36.4 ms for a single yes/no.
Quickstart
bash
hf download startlux-models/StartLux-Decision-v1-35B-A3B --local-dir StartLux-Decision-v1-35B-A3B
cd StartLux-Decision-v1-35B-A3B
pip install -r requirements.txt
python -m startlux_decision.check .
python -m startlux_decision.server --model . --port 8090Watch out
Card reviewed 2026-10-17; the Decision Index, JevBench, Intern-Decision and latency figures are self-reported on the card's own results panel, and the 35B-A3B still trails the family's 27B by 2.3 points on Decision Index 0.2.1 (61.55 versus 63.88). The cc-by-nc-4.0 licence excludes commercial use, and the MLX and GGUF paths cover only text and the dense sizes.
Questions about this entry
Short answers drawn from the linked project or page.
The 35B-A3B mixture-of-experts member of StartLux's Decision v1 family. It answers typed questions about a state — pick one of several options, yes/no or a rating — returning a probability for every option using the TypeSafe /v1/systemone format.
The card says requests and responses use the TypeSafe /v1/systemone format so clients written for Jev work unchanged. On the card's comparison table it reports 210/231 JevBench public versus Jev 1.13's 199, and Decision Index 0.2.1 61.55 versus Jev's 57.91.
The card links 0.8B, 2B, 4B, 9B, 27B and this 35B-A3B variant. The card notes the 35B-A3B answers in half the time of the 27B and has the family's highest JevBench and Intern-Decision numbers.
One 80 GB GPU is enough: about 65 GiB of GPU memory in bf16, peaking around 71 GiB with CUDA graphs and a batch of 256 requests, and a 72 GB download. Its own package checks for flash-linear-attention and causal-conv1d fast kernels.
52.5 ms end to end over HTTP for one request with three questions and 36.4 ms for one yes/no question on one H200 in bf16, with Decision Index 0.2.1 of 61.55, 210/231 on public JevBench and 92.29 Intern-Decision average over seven suites.
Yes: the weights include a vision tower, images can be passed as images=[...] in Python or a JSON array over HTTP with <image> markers in the state, and the model's native context is 262,144 tokens, read once with questions branching off it.
cc-by-nc-4.0, so non-commercial only. Download with hf download, install requirements.txt, verify fast kernels with python -m startlux_decision.check, then run python -m startlux_decision.server --model . --port 8090 and POST to /v1/systemone.
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 27B decision model from StartLux Labs that takes text, JSON or images up to 262,144 tokens and returns a probability per option through the TypeSafe /v1/systemone format, reporting Decision Index 0.2.1 63.88 and 208 of 231 on JevBench public.
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
GGUF quantizations of StartLux Labs' 0.8B Instant decision model for llama.cpp: Q8_0 (0.81 GB, DecisionPivot 9.7) and Q4_K_M (0.53 GB, 8.7), plus a 205 MB F16 vision encoder, against 9.5 for BF16 on vLLM.
Jev Models & Reimplementations#community#open-models#local
Community
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.
Jev Models & Reimplementations#community#open-models#local
Community
StartLux Labs' 4B Instant decision model returns a probability for every option of every typed question from one forward pass with no generated text; it self-reports Decision Index 0.3 public 67.88 and DecisionPivot pair accuracy 41.5.
Jev Models & Reimplementations#community#open-models#local
Community
A 27B decision model from torchcast-ai that answers choice, noul and score questions together in one forward pass over the TypeSafe-compatible /v1/systemone format; the card reports Decision Index 0.2.1 of 65.00, +7.09 over Jev 1.13, at 33 ms per yes/no question.
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.