Jev directory
CommunityJev Models & Reimplementations
Bonsai-JEV
A calibrated image-capable 27B decision model that fits an 8 GB GPU: the JEV LoRA-and-head recipe trained on Prism ML's ternary Bonsai 2, returning typed probabilities with no generated text.
- Reading time
- 3 min
- Published by
- Community
- Author
- dakebakery
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- macOS
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Adds the JEV recipe (a 108.8M-param LoRA plus a 123k decision head) to Ternary Bonsai 2 27B at 2.13 bits/weight; the download is about 8.7 GB.
- Runs on an 8 GB GPU with PTQ1_0 at ~7.1 GiB peak including vision and 4k context, and on 16 GB Apple Silicon via MLX or a Radeon 780M iGPU via Vulkan.
- Self-reported KL mean 0.0163 on the jev-distill-corpus-v3 test_set_30k, ahead of JEV-9B's published 0.019 and JEV-27B's 0.017.
- VL-RewardBench 78.0 overall and 84.9 on hallucination, a tie with JEV-27B-VL's 78.3; ERQA 52.0 on 400 robot-scene questions.
- Decision Index public vision suite 67.56; the text suite was not run, and with the adapter off the same files serve as Ternary Bonsai 2 byte for byte.
Watch out
Card reviewed 2026-10-17; all numbers are self-reported through the author's own deployment path, the Decision Index text suite was never run, a 1,247-pair difference of ~2.4 points is called a tie, the 8 GB GPU result is a simulation on an RTX 4070 capped at 7.5 GiB, and the project is independent of AutoTrust, TypeSafe and Prism ML.
Questions about this entry
Short answers drawn from the linked project or page.
A calibrated, image-capable 27B decision model that fits on an 8 GB GPU. It applies the JEV recipe (a LoRA plus a 24-slot decision head, distilled to soft labels) to Prism ML's ternary Bonsai 2 27B.
Typed yes/no, 0-5 score and 2-16-way choice questions about a state that can include images. One forward pass returns calibrated probabilities over the options, and no text is generated.
On the same test rows it reports KL mean 0.0163 against JEV-27B's published 0.017, choice top-1 agreement with Jev of 91.1% against 90.5%, and VL-RewardBench 78.0 against 78.3, called a tie at 1,247 pairs.
An 8 GB GPU with the PTQ1_0 build at ~7.1 GiB peak including vision and 4k context (simulated on a capped RTX 4070), a 12 GB RTX 4070, 16 GB Apple Silicon via MLX on a Mac mini M4, or a Radeon 780M iGPU via Vulkan.
About 8.7 GB at PQ2_0 (2.13 bits/weight) or about 7.4 GB at PTQ1_0, versus ~54-56 GB for JEV-27B and JEV-27B-VL in bf16. The ternary base weights are included byte-identical with the frozen base.
Yes. With the adapter switched off, the server is Ternary Bonsai 2 itself, byte for byte, so the same weights serve System 2 text generation and System 1 typed decisions depending on the adapter.
Everything is measured through the author's own deployment path, the Decision Index text suite was not run, ERQA carries about plus or minus 5 points at 400 questions, and the project is independent of AutoTrust, TypeSafe and Prism ML.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsX: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
More Jev resources like this
A Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
A LoRA adapter for Qwen3-VL-4B-Instruct that scores supplied visual answer candidates instead of writing prose: macro accuracy over four choice benchmarks rises from 0.706 to 0.761, Apache-2.0.
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 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
A 4-bit AWQ plus GPTQ build of AutoTrust's JEV-35B from AI Outfitters: INT4 weights with FP8 and INT8 activations, calibrated on all 256 experts, and served by vLLM on a single 24 GB GPU.
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
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.