Jev directory
CommunityJev Models & Reimplementations
StartLux-Decision v1.1 4B Instant
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.
- 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 67.88 and DecisionPivot pair accuracy 41.5, below the 4B AutoThink sibling's 68.21 and 59.9.
- config.json declares jev_prompt_attention bidirectional; a causal engine silently serves a different model without an error.
- Stage 1 fine-tuned all weights on about 2.7 million decision examples; stage 2 trained and merged an adapter on about 1.7 million more.
- Released weights are the uniform average of four runs differing only in random seeds; weights are CC BY-NC 4.0.
- vLLM BF16 median read-out: 31 ms at 256 tokens and 142 ms at 8,192; MLX builds run on an Apple M5 Pro with 64 GB of memory.
Watch out
Card reviewed 2026-10-17; the Decision Index and DecisionPivot rows are StartLux's own measured runs, which the card says were used to compare training recipes and seed averages, so they reflect model selection rather than independent tests. Weights are CC BY-NC 4.0, and the card warns that an engine reading the prompt causally serves a different model.
Questions about this entry
Short answers drawn from the linked project or page.
A 4B decision model from StartLux Labs. You give it a state (text or JSON, with up to 8 images) and typed questions - choice, noul yes/no or score - and it returns a probability for every option of every question from one forward pass, generating no text.
Each question is read as its own prompt that holds the whole state, and at the last prompt position the logits of the option letters are divided by a question-type temperature and normalised over the options. There is no output text: the answer is the distribution itself.
In one of four full-attention layers, prompt tokens also attend to later prompt tokens so the state is read with the question in view, while linear-attention layers stay causal. config.json declares this with jev_prompt_attention set to bidirectional, and a causally reading engine serves a different model without an error.
Stage 1 fine-tuned all weights on about 2.7 million decision examples; stage 2 trained an adapter with bidirectional prompt attention on about 1.7 million further examples and merged it. The release is the uniform average of four such runs that differ only in random seeds.
The card reports Decision Index 0.3 public 67.88 and DecisionPivot pair accuracy 41.5, against 68.21 and 59.9 for the 4B AutoThink sibling. The card says its scores are its own runs used for model selection, not independent test results.
Clone StartLuxLabs/StartLux-Decision, pip install the vllm extra, download the checkpoint and startlux-decide serve --launch-vllm. vLLM on one NVIDIA H200 measured 31 ms median at 256 tokens; MLX 8-bit and 4-bit builds run on an Apple M5 Pro with 64 GB.
CC BY-NC 4.0: free for research and other non-commercial use, with commercial use requiring a separate licence from StartLux Labs. The inference code in the StartLuxLabs repository is distributed separately from the weights.
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
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'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.
Jev Models & Reimplementations#community#open-models#multimodal
Community
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
A local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.
Jev Models & Reimplementations#community#rust#open-models
Community
ollaya-dev/ollayaRust
ollaya-dev/ollayaRustFrom 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.