Jev directory
CommunityJev Models & Reimplementations
StartLux-Decision v1.1 2B Instant GGUF
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.
- 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
- Q8_0 file is 2.01 GB and scores 23.7 DecisionPivot on llama.cpp; Q4_K_M is 1.27 GB and scores 21.9; BF16 on vLLM scores 23.5.
- Files hold the text decoder only, without the BF16 checkpoint's multi-token-prediction layer; images need the separate 668 MB F16 mmproj file.
- Instant models read prompts bidirectionally, so llama-server must evaluate every prompt alone: one slot, no context checkpoints, micro-batch 16384.
- The card warns that a stock server's default batching cost the 4B Instant sibling 7.7 DecisionPivot points, and a causal read loses 27 of 48.
- Weights are CC BY-NC 4.0 and Q8_0 is the recommended file; Q4_K_M lost up to 3.8 DecisionPivot points; tested with llama.cpp b11549.
Quickstart
bash
llama-server -m ~/models/StartLux-Decision-v1.1-2B-Instant-GGUF/StartLux-Decision-v1.1-2B-Instant-Q8_0.gguf -ngl 99 -c 20480 \
--parallel 1 --ctx-checkpoints 0 -b 16384 -ub 16384 --port 8081 &
STARTLUX_LLAMACPP_BIDIR=server startlux-decide serve --model ~/models/StartLux-Decision-v1.1-2B-Instant \
--backend llamacpp --llamacpp-url http://127.0.0.1:8081 --port 8090Watch out
Card reviewed 2026-10-17; the DecisionPivot figures are StartLux's own llama.cpp b11549 measurements from its technical report rather than independent runs, and the vision encoder is a separate download. The weights are CC BY-NC 4.0, so commercial use needs a separate licence, and the GGUF files drop the multi-token-prediction layer present in the BF16 checkpoint.
Questions about this entry
Short answers drawn from the linked project or page.
A llama.cpp GGUF package for StartLux-Decision-v1.1-2B-Instant, a 2B decision model from StartLux Labs. Given a state (text or JSON) and typed questions (choice, noul or score), it returns a probability for every option of every question from one forward pass without generating text.
The card says requests and responses use the TypeSafe /v1/systemone format: you send a state and typed questions, and the model returns probabilities over your listed options instead of generated text, which is the same shape as the System One contract behind Jev.
Q8_0 (2.01 GB) and Q4_K_M (1.27 GB) text-decoder builds, plus an F16 vision encoder mmproj file of 668 MB. Q8_0 is the recommended file; Q4_K_M lost up to 3.8 DecisionPivot points across the sizes, and both files omit the multi-token-prediction layer of the BF16 checkpoint.
Download the quant and also the tokenizer and configs from the BF16 repository, install the StartLux-Decision package, then serve through llama-server with the startlux-decide wrapper. Because prompts are read bidirectionally, the stock server needs one slot, no context checkpoints and a full-length micro-batch (-b 16384 -ub 16384). Tested with llama.cpp b11549.
DecisionPivot is the mean chance-corrected accuracy over seven scorer layers on 7,800 core items. The card reports 23.7 for Q8_0 on llama.cpp and 23.5 for BF16 on vLLM; these are StartLux's own measurements from its technical report, not independent runs.
StartLux Labs published the weights under CC BY-NC 4.0, which allows research and other non-commercial use, while the code in the StartLuxLabs/StartLux-Decision repository has separate terms. Commercial use of the weights needs a licence from the publisher.
It is the 2B Instant build; the family also ships 0.8B, 4B, 9B and 27B models in Instant and AutoThink variants. AutoThink models write a thought only when confidence is low, while this Instant model always answers from a single read-out.
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' 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
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
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
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
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.