Jev directory
CommunityJev Models & Reimplementations
StartLux-Decision v1 9B Q8_0 GGUF
The Q8_0 8-bit GGUF build of StartLux-Decision-v1-9B, 9.53 GB for llama.cpp, with the card reporting 100.0% same decisions as the original weights and 200 of 231 correct (86.6%) on JevBench public.
- Reading time
- 2 min
- Published by
- Community
- Author
- StartLux Labs
- Last verified
- Oct 2026
- Languages
- Shell
- Capabilities
- Self-hostedRuns locally
Highlights
- Q8_0 build of StartLux-Decision-v1-9B: 9.53 GB and 100.0% same decisions as the original on JevBench public.
- Answers 200 of 231 JevBench public items (86.6%); Q4_K_M agrees on 98.3% of items and BF16 on 99.6%.
- Files: Q8_0 9.53 GB recommended, Q4_K_M 5.63 GB smallest, BF16 17.92 GB the original weights unchanged.
- Run llama-server with -ngl 99 for GPU offload, then startlux_decision.gguf_server points at it on port 8090.
- The TypeSafe /v1/systemone format matches the original model, so Jev-shaped clients work unchanged.
Quickstart
bash
hf download startlux-models/StartLux-Decision-v1-9B-Q8_0-GGUF --local-dir StartLux-Decision-v1-9B-Q8_0-GGUF
cd StartLux-Decision-v1-9B-Q8_0-GGUF
pip install -r requirements.txt
llama-server -m StartLux-Decision-9B-Q8_0.gguf -ngl 99 -c 16384 --parallel 4 --port 8081 &
python -m startlux_decision.gguf_server --model-dir . --llama http://127.0.0.1:8081 --port 8090Watch out
Card reviewed 2026-10-17; the agreement percentages and JevBench counts are the publisher's own measurements, and the weights remain CC BY-NC 4.0 while the inference code is Apache-2.0. Plain llama.cpp chat does not produce typed decisions; the readout lives in gguf_server.py.
Questions about this entry
Short answers drawn from the linked project or page.
The Q8_0 8-bit GGUF conversion of StartLux-Decision-v1-9B, 9.53 GB for llama.cpp. The bundled decision server produces the same typed decisions, with a probability for every option, as the original model.
The card measures 100.0% same decisions as the original on the 231 public JevBench items, with 200 correct (86.6%). Q4_K_M agrees on 98.3% of items at 5.63 GB, and BF16 on 99.6% at 17.92 GB.
Q8_0 at 9.53 GB is marked recommended, Q4_K_M at 5.63 GB is marked smallest, and the BF16 file at 17.92 GB is the original weights unchanged. All are llama.cpp quantizations of the released BF16.
Download the repository, pip install -r requirements.txt, run llama-server with StartLux-Decision-9B-Q8_0.gguf and then python -m startlux_decision.gguf_server --model-dir . --llama http://127.0.0.1:8081 --port 8090; llama.cpp build b10454 or newer is required.
In the example command -ngl 99 puts every layer on the GPU (CUDA or Metal); leave it out on a CPU-only machine. A CPU build of torch is enough for the wrapper, and llama.cpp must be recent enough for this model.
No. The file holds the text decoder only, and the option-letter readout plus the per-type temperatures are implemented in gguf_server.py; serving the weights from llama.cpp alone gives text continuation, not typed answers.
CC BY-NC 4.0 for the weights: non-commercial use with attribution, with commercial use requiring a separate licence from StartLux Labs; the startlux_decision inference code is Apache-2.0.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
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 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
The Q8_0 8-bit GGUF build of StartLux-Decision-v1-27B, 28.60 GB for llama.cpp, served through a decision wrapper that the card measures at 100% same decisions as the original and 209 of 231 on JevBench public.
Jev Models & Reimplementations#community#open-models#quantized
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
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 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
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/ollayaRustA community Q8_0 GGUF of Mapika/decider-0.8b, the smallest System One typed-decision model: one forward pass returns calibrated Choice, Score and noul answers, reporting 0.405 on JevBench hard-public at 920 MB resident.
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.