CommunityJev Models & Reimplementations
Jev-Style Qwen3.5-2B MLX
An Apple Silicon MLX bf16 build of the v1 Jev-Style decision model: 77 ms decisions on an M1 Max with calibrated choice, bool and score probabilities, Apache-2.0 weights.
- Published by
- Community
- Author
- chaoliangUNSW
- Last verified
- Oct 2026
Highlights
- MLX bf16 build of the v1 Jev-Style model on Qwen3.5-2B-Base, with a native jev_style_mlx.py client for Apple Silicon.
- 82.3% accuracy and ECE 0.017 on 1,500 held-out examples across five decision tasks, against the base model's 65.9% and 0.065.
- 77 ms per decision on an M1 Max under MLX bf16, the card says at no added cost over the unmodified base model.
- Reads choice, bool and score primitives in one prefill pass; up to 26 options natively and 20 through a server's top_logprobs.
- LM Studio's MLX engine does not return token log-probs, so the card points to the GGUF build for probability readouts.
Watch out
LM Studio's MLX engine cannot read the probabilities, so on macOS you either run jev_style_mlx.py or switch to the GGUF sibling. The model needs the card's exact prompt format with 'Answer:' as the final line, weights are Apache-2.0 like the Qwen base, and AG News training data carries research/non-commercial terms.
Questions about this entry
Short answers drawn from the linked project or page.
decide returns per-option probabilities for a choice question, decide_bool returns a single probability that a proposition is true (0.976 in the card's entailment example), and decide_score returns both the expected level index and the full distribution over ordered levels.
LM Studio's MLX engine does not return token log-probs, so probabilities cannot be read inside it. Use the GGUF sibling with jev_style_client.py, or run jev_style_mlx.py directly with MLX on the command line.
pip install mlx-lm, then run python jev_style_mlx.py. The JevStyle class downloads this repository on first use, and jev.decide(state, question, options) returns a list of option and probability pairs.
82.3% on 1,500 held-out examples across five decision tasks, against 65.9% for Qwen3.5-2B-Base zero-shot, with ECE 0.017 against 0.065, NLL 0.418 and Brier 0.242. The temperature is folded into the final RMSNorm weight.
Partly: on emotion classification and RTE, which were never in training, calibration error halved from 0.155 to 0.075 at unchanged accuracy of 64.5%. The card describes this as better but not perfect calibration.
Qwen/Qwen3.5-2B-Base, fine-tuned with LoRA rank 16 on all linear layers using five English task families — SST-2, MNLI, AG News, BoolQ and SST-5 — totalling 22k examples converted to typed decisions.
The MLX release is bf16, the card's full-precision reference. Its GGUF sibling reports Q8_0 making the same decision as bf16 on 99.4% of examples and Q4_K_M at 1.3 GB keeping 82.4% accuracy, so the quants trade exactness for size.
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
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 NewsHacker 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 NewsRelated terms
Glossary definitions related to this entry.
More Jev resources like this
Ruiruiz30's Apple Silicon MLX conversion of Jev-Omni: 4-bit affine language weights on a Mac mini M4 16GB, about 7 GB peak Metal memory and roughly 963 ms text decisions, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
Jev Models & Reimplementations#community#open-models#local
Community
akhilaaa3's merged multimodal decision classifier on Gemma 4 12B IT: probabilities for text, image, audio or video options, 87.57% on DecisionBench Medium and MMAU 63.10%, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
Reza2kn's llama.cpp Q4_K_M build of Jev-Omni: a 6.87 GiB multimodal decision backbone plus projector and FP32 head, measured at 8.98 GiB VRAM on an RTX 5080 Laptop GPU.
Jev Models & Reimplementations#community#open-models#local
Community
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.