Jev directory
CommunityJev Models & Reimplementations
NanoJev on ModelScope
A community re-upload on ModelScope of NanoJev, the 0.6B Qwen3-based parallel decision model: full probability distributions over 2-255 dynamic choices in one batched forward pass with zero output-token decoding, approval-gated and license-free.
- Reading time
- 2 min
- Published by
- Community
- Author
- Tekdycaml
- Last verified
- Oct 2026
Highlights
- 0.6B parallel decision model on a Qwen3-0.6B backbone with structured decision heads and zero output-token decoding.
- Choice supports 2-255 dynamic candidates, boolean returns a probability, and score spans 2-10 ordered levels with an expected score.
- States and questions batch into one backbone forward; the readme's measured service ran 6 states, 18 questions, and 44 candidate paths in one pass.
- Re-upload of the Hugging Face NanoJev under Tekdycaml, with variants for a 50x50 maze, a 12x12 Snake, and calibrated-decision experiments.
- The page sets ApprovalMode 1 and ProtectedMode 2, lists en and zh, and showed 526 downloads with eight 2.38 GB F32 safetensors files.
Watch out
The ModelScope page lists no licence, an empty License field, and ApprovalMode 1 suggesting approved access, and its description says it is a re-upload of the Hugging Face NanoJev that readers should consult themselves, so terms come from the original project. Training specifics for the checkpoint itself are not published on this page.
Questions about this entry
Short answers drawn from the linked project or page.
A 0.6B parallel decision model, described in its readme as a nano replica of Jev, built on a Qwen3-0.6B backbone with structured decision heads. It returns complete probability distributions over dynamic candidates with zero output-token decoding.
Choice questions with 2 to 255 dynamic candidates and a probability for each, boolean questions returning the probability that a proposition is true, and score questions with 2 to 10 ordered levels, their distribution, and an expected score.
The readme's running-service measurement evaluates multiple states and questions in one batched backbone forward: 6 states, 18 questions, and 44 candidate paths in a single pass, with full distributions for greedy selection or probability sampling.
Specialized variants for the game showcases: local_atomic_seed17 for safety judgments in a 50x50 maze, games_gold_seed17 for dynamic action choices in a 12x12 Snake, games_api_seed17 for full-map game-question comparison, and events cross-entropy and Brier controls.
Tekdycaml's description says it is a re-upload of the Hugging Face NanoJev, and the Chinese name records that it was moved from fredzzh/NanoJev. The readme links source code at github.com/TianyuCodings/NanoJev and a dataset on Hugging Face.
No licence appears in the ModelScope metadata; the License field is empty. The page also shows ApprovalMode 1, suggesting approved access, so check the original Hugging Face project for the terms that actually apply.
The page lists eight safetensors files of 2,385,039,280 bytes each in F32, including the base best.safetensors, a stage1 copy, and six variant checkpoints, plus a model_size metadata figure of about 4.77 GB.
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
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
A LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
The original OSS Jev Qwen2.5-0.5B decision checkpoint, trained 2,000 steps at learning rate 2e-05; the 10k sibling card warns that this overwrite tip may be the worse noul-fix variant and recommends sized siblings.
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 Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
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.