Jev directory
CommunityJev Models & Reimplementations
Gemma 4 12B Unified System One GGUF
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.
- Reading time
- 3 min
- Published by
- Community
- Author
- knowlet
- Last verified
- Oct 2026
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- GGUF Q8_0 conversion of the merged BF16 checkpoint, with a native S1 decision runtime and separately fitted GGUF temperature.
- BoolQ specialist: one epoch of rank-8 LoRA on 2,048 English examples, merged into the full multimodal model; not full-parameter training.
- Two files: s1-boolq-Q8_0.gguf at 12,669,647,264 bytes and mmproj-s1-boolq-f16.gguf at 122,031,744 bytes, 12,791,679,008 bytes total.
- Twelve native execution checks passed on Apple M1 Max, 64 GiB, covering text, image, audio and an 8,481-token request crossing the batch boundary.
- Choice, Noul and Score are supported schemas, but training establishes evidence only for passage-grounded English yes/no decisions.
Watch out
Card reviewed 2026-10-17; the twelve native execution checks and the 8 requests with 29 typed answers are functional smoke tests, not accuracy or calibration measurements, and the card says matching tokens and positions does not imply numerical parity with BF16 or MLX. The runtime targets Apple Silicon/macOS and needs the custom native S1 runner, since a generic chat-generation command does not implement this decision interface. The fetched card text ends before the full build steps, so no run command is quoted here.
Questions about this entry
Short answers drawn from the linked project or page.
A complete GGUF conversion by knowlet of the trained, merged BF16 checkpoint, with a native S1 decision runtime and an independently fitted GGUF temperature. The language model uses Q8_0 storage and a separate F16 mmproj file holds the native image and audio projections.
As a BoolQ specialist: one epoch of rank-8 LoRA (alpha 16, q/k/v/o, cross-entropy, learning rate 2e-5, seed 42) on 2,048 English examples, merged into the complete multimodal model before conversion. The card stresses it is not a full-parameter training run.
Choice, Noul and Score are supported interface schemas, but the card states the training establishes evidence only for passage-grounded English yes/no decisions, so the other schemas should be treated as unproven for this checkpoint.
Twelve native execution checks passed on an Apple M1 Max with 64 GiB, covering text, image, audio, mixed media, 16 questions, a 1,299-token request and an 8,481-token request whose answer slots crossed the 8,192-token batch boundary. These are functional checks, not benchmark scores.
The implemented Python adapter and native helper score all unfilled answer slots in their original question order, never sampling or inserting earlier answers. The native graph computes the full-vocabulary head, then selects legal candidate logits and normalizes them with the fitted GGUF temperature.
The instructions target Apple Silicon and macOS with Git, CMake, a C++17 compiler and uv installed; the measured host had 64 GiB unified memory. The card notes the weight-file size alone is not the runtime memory requirement.
The card declares apache-2.0 and the original base is google/gemma-4-12B-it at revision 707f0a3b; the trained BF16 weights come from knowlet/Gemma-4-12B-Unified-System-One. Datasets metadata points to google/boolq, and the core GGUF files are pinned to an immutable revision.
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
An 8-bit MLX conversion by SirSahOl of chaoliangUNSW/Jev-Style-0.8B-Decision-v3 built for Apple Silicon native GPU inference: about 990 MB of active VRAM, a 262,144-token context and an average of 8.25 bits per weight.
Jev Models & Reimplementations#community#open-models#local
Community
Perplexity AI's 27B decision model fine-tuned from Qwen/Qwen3.8-27B, producing option probabilities through a custom head instead of generated text; the card reports 85.71% overall across 11 benchmarks against Jev's 84.51%.
Jev Models & Reimplementations#community#open-models#multimodal
Community
Fluid Inference's fixed-shape Core ML conversion of the classification path of Fastino's GLiNER2.5-Decide for Apple silicon: the DeBERTa-v3-large encoder and shared label classifier score up to four decision heads per call, in fp16, 8-bit or 6-bit packages.
Jev Models & Reimplementations#community#open-models#local
Community
An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonA 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.
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
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.