Jev directory
CommunityJev Models & Reimplementations
Bekko System One v0 68M
The 68M member of hotchpotch's Bekko System One v0 family: a compact English encoder for choice, noul and score decisions that returns candidate probabilities in parallel, with a ~196 MB browser ONNX export and no text generation.
- Reading time
- 2 min
- Published by
- Community
- Author
- hotchpotch
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- Browser
- Capabilities
- Self-hostedRuns locally
Highlights
- 68M-parameter English encoder for Choice, Noul and Score decisions; model.predict() returns candidate probabilities, a selected option or a rubric score.
- Part of the v0 family (17M, 68M, 400M) built on Ettin rerankers with a shared-prefix encoder and task-specific heads, evaluating candidates in parallel.
- Encodes instructions and state once per unique tokenized prefix in a microbatch and reuses the computation across candidates.
- Browser ONNX file is about 196 MB; exports quantize the token lookup table to row-wise INT8 while transformer blocks and task heads stay FP32.
- The card itself warns that v0 falls far behind Jev 1.13 on S1MB's generalization benchmarks and that strong task results do not establish broad generalization.
Quickstart
bash
pip install 'torch>=2.10,<2.11' 'transformers==5.17.0' 'sentence-transformers==6.1.0' 'safetensors>=0.7' 'tqdm>=4.67'Watch out
Card reviewed 2026-10-17; the card explicitly frames v0 as limited — it says the family falls far behind Jev 1.13 on S1MB benchmarks designed to measure generalization and that training data includes dataset families represented in the evaluation, so task scores are not evidence of broad generalization. No licence is stated, and parameter counts are rounded.
Questions about this entry
Short answers drawn from the linked project or page.
The 68M variant of Bekko System One v0, hotchpotch's family of compact English encoder models for Choice, Noul and Score decisions. Supply instructions, application state and candidate definitions, and model.predict() returns candidate probabilities, a selected option or a numeric rubric score.
The card places Bekko System One in the same category as TypeSafe AI's Jev, but explicitly warns the v0 models fall far behind Jev 1.13 on S1MB's generalization benchmarks; the v0 designation reflects that limited generalization.
Choice selects among named alternatives, Noul estimates the probability of an authored binary condition, and Score returns the expected value of a numeric rubric. All work over supplied candidates in one encoder pass, with no generated text.
17M, 68M and 400M parameters. This 68M checkpoint has 42M active parameters excluding lookup-only embeddings, and its browser ONNX file is about 196 MB (the 17M is about 29 MB and the 400M 1.43 GB).
Install torch, transformers, sentence-transformers, safetensors and tqdm, then load BekkoSentenceTransformer with get_class_from_dynamic_module from the repo using trust_remote_code=True on CPU or CUDA, pass state_json plus decisions, and read selected_id and probabilities.
The browser demo linked from the card offers the 17M and 68M variants with CPU or WebGPU execution. Browser exports quantize the token lookup table to row-wise INT8 while transformer blocks and task heads remain FP32, so they are not fully INT8 models.
No licence is stated on the card. The encoder derives from cross-encoder/ettin-reranker-68m-v1, and Bekko System One builds on Ettin rerankers with a shared-prefix encoder and task-specific heads; the training dataset is published separately as bekko-system-one-dataset-v0.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Text-to-screenshot (Malay Vasa)
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Google's MediaPipe Decision Maker, tested in the browser
More Jev resources like this
The smallest Bekko System One v0 model: a 17M-parameter English encoder (3.90M excluding lookup-only embeddings, about 29 MB as browser ONNX) that answers choice, noul and score decisions without generating text.
Jev Models & Reimplementations#community#open-models#local
Community
FRIDA-Decisions answers Russian text with structured decisions in a single encoder pass — choice, ordinal score, yes/no or ranking — and scores 0.893 on the 735-item razvilka benchmark, effectively level with TypeSafe Jev's 0.897 (McNemar p = 0.84).
Jev Models & Reimplementations#community#open-models#russian
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
The 9B entry in vLLM Semantic Router's Decision 2.0 family: a 7.94B decision model with a 16,384-token context that answers choice, yes/no and score questions in one pass, reporting 68.1 JevArena and 18.4 ms median single-question latency.
Jev Models & Reimplementations#community#open-models#local
Community
The 4B model in vLLM Semantic Router's Decision 2.0 family: a 4.21B fine-tune of Qwen3.5-4B-Base that answers choice, yes/no and score questions in one forward pass, with a self-reported JevArena 63.6 and a 12.9 ms median per request.
Jev Models & Reimplementations#community#open-models#local
Community
The 0.6B model in vLLM Semantic Router's Decision 2.0 family: a 0.60B fine-tune of Qwen3-0.6B-Base with an 8,192-token context that returns probabilities for choice, yes/no and score questions from one forward pass.
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.