Jev directory
CommunityJev Models & Reimplementations
Jev-Chess 1 Small 6.5M
A 6.5M-parameter independent chess policy model built from Jev's public description: it returns a calibrated distribution over legal moves in one forward pass, with a self-reported held-out top-1 of 36.5%, ECE 0.036 and Elo 1239 plus or minus 51.
- Reading time
- 2 min
- Published by
- Community
- Author
- omerfaral
- Last verified
- Oct 2026
Used inJev for Gaming
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 6.5M parameters (d=256, 8 layers, 8 heads), trained 30,000 steps on 5,000,000 Lichess positions labeled with Stockfish's best move.
- Self-reported held-out top-1 of 36.5% and calibration ECE of 0.036; Elo 1239 plus or minus 51 over 200 games.
- The Elo was measured against Stockfish UCI_Elo 1320 and 1400 on the CCRL scale at 60s+0.6s.
- Independent reimplementation from Jev's public description, not TypeSafe's weights; a pure policy network with no search.
- Runs as a UCI engine or a Python call, returning move-probability pairs from one parallel forward pass.
Quickstart
python
from uci import load; from jev import predict; import chess
model, _ = load("omerfaral/jev-chess-1-small-6.5M")
print(predict(model, chess.Board())[:3]) # [(move, prob), ...]Watch out
Card reviewed 2026-10-17; the 36.5% top-1, 0.036 ECE and 1239 plus or minus 51 Elo are self-reported from the author's own evaluation. The card states plainly that TypeSafe has not published Jev's internals, so this is an independent model built from Jev's public description rather than a port of the official weights. Licence is CC0-1.0.
Questions about this entry
Short answers drawn from the linked project or page.
A 6.5M-parameter model that predicts the next best chess move. The board state goes in and a single parallel forward pass returns a calibrated probability distribution over the legal moves, Jev-style.
TypeSafe has not published Jev's internals, so the author built it from Jev's public description: the state goes in and one forward pass returns a Choice-style distribution. Chess has at most 218 legal moves where Jev's Choice takes up to 255.
6.5M parameters with d=256, 8 layers and 8 heads, trained for 30,000 steps on 5,000,000 Lichess positions labeled with Stockfish's best move (PV1). It is a pure policy network with no search.
Held-out top-1 is 36.5%, calibration ECE is 0.036, and its Elo is 1239 plus or minus 51 on the CCRL scale over 200 games at 60s+0.6s against Stockfish UCI_Elo 1320 and 1400, all self-reported on the card.
With the card's Python snippet: load the model and call predict on a chess board to get move-probability pairs, or run it as a UCI engine with python uci.py; it needs torch, chess, safetensors and huggingface_hub.
CC0-1.0, so it is effectively public domain. The training data is the Lichess/chess-position-evaluations dataset, and the repository is a safetensors checkpoint with no commercial restrictions stated on the card.
It plays one move per forward pass with no search, and its 36.5% top-1 means most of its first-choice moves differ from Stockfish's best move, though the calibrated distribution still ranks candidates; its 1239 Elo is modest.
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
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
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
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
Jev Models & Reimplementations#community#open-models#local
Community
A bilingual Turkish and English 4B Jev-class decision model from Berkay Hasret: a roughly 130 MB LoRA adapter on Qwen/Qwen3.5-4B that answers up to 16 typed questions per text in one pass and reports 0.774 JevBench macro accuracy.
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.