Jev directory
CommunityJev Models & Reimplementations
Laya on Ollama
Convai Innovations' 421M non-autoregressive decision model on Ollama: fine-tuned from ModernBERT-large, Apache-2.0, answering one question in under 10 ms on an M5 Max with 0.060 in-task calibration error.
- Reading time
- 2 min
- Published by
- Community
- Author
- Convai Innovations
- Last verified
- Oct 2026
- Platforms
- CLI
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 421M parameters, fine-tuned from ModernBERT-large; one question can be answered in under 10 ms on an M5 Max.
- Trained with RLCD against strictly proper scoring rules, so honest probabilities maximize reward; in-task calibration error (ECE) is 0.060.
- Scores every option of every question in a single forward pass with no decoder, so there is no text to parse and nothing to hallucinate.
- Apache 2.0 licence with 7,328 downloads and four checkpoints, including multilingual and typed-decisions MLX builds at 678-846 MB.
- Shipped checkpoint is fine-tuned on human-annotated email triage, content safety, intent routing, and fact verification.
Watch out
Apache-2.0 weights, but the shipped checkpoint targets a fixed set of workflows, and every option shares a fixed token budget inside the decision head, so the page advises keeping a single choice question to roughly 20 options. Decision models were not yet in the Ollama CLI or client libraries, so calls go through /v1/systemone or the TypeSafe SDK.
Questions about this entry
Short answers drawn from the linked project or page.
A 421M-parameter decision model from Convai Innovations, fine-tuned from ModernBERT-large, that returns typed answers with calibrated probabilities in a single forward pass and never generates text.
Run ollama pull laya, which requires Ollama 0.40.0 or later. Requests go to the /v1/systemone endpoint, because decision models were not yet in the Ollama CLI or the Python and JavaScript libraries at capture.
Apache 2.0. The Ollama library lists four checkpoints: laya:latest at 846 MB, plus MLX builds for the 322M multilingual, 421M English, and 421M typed-decisions variants at 512 to 1K context.
The model was trained with RLCD against strictly proper scoring rules, so reporting honest probabilities is the only way to maximize reward; its reported in-task calibration error (ECE) is 0.060. Per-task accuracy is not broken out on the page.
Real, human-annotated workflows: email triage covering spam, phishing, and department routing; content safety; intent routing; and fact verification. The page presents these as the checkpoint's training domains rather than a benchmark suite.
Up to 64 named questions about the same text in one call, each returning an entry under answers. The page notes they share a single forward pass with no autoregressive decoding, and answers come back in the order asked.
Confidence runs from 0 to 1 and reflects how concentrated the probabilities are, not the probability the answer is right. The page warns that large label sets narrow what each option looks like and suggests staying near 20 options per choice question.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Both Jev 1.13 and 1.14 score 29 of 30 on 30 Turkish buyer and seller messages
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 NewsMore Jev resources like this
ActBro's 2B calibrated System 1 decision model on Ollama, built on OpenBMB's MiniCPM5-2B: 78.79% on JevBench among sub-2B open-weight models, 86.41% on the 10-source decision-v7 set, and one output token per decision via /v1/systemone.
Jev Models & Reimplementations#community#open-models#local
Community
The Laya family mirror on ModelScope: a 421M non-autoregressive System 1 decision model, Apache-2.0, with a pip-installable package, more than 100 languages claimed at about 33 ms per pass, and 16,713 downloads.
Jev Models & Reimplementations#community#open-models#local
Community
A 0.6B non-autoregressive decision model on Qwen3-0.6B that returns calibrated Choice, Score and Noul distributions in one forward pass, scoring 95.8% on its 22,110-item held-out test split.
Jev Models & Reimplementations#community#open-models#local
Community
Fastino's 287M multilingual classification model in the GLiNER2.5 family: pass any label set at call time and get labels back in a single forward pass, with no prompt template and no generated tokens. It averages 56.7% exact-match on the Fast Decisions suite.
Jev Models & Reimplementations#community#open-models#local
Community
Cloudflare's 9B multimodal decision model on Ollama: text, JSON, or images in one non-autoregressive forward pass, Apache-2.0, and the lowest measured latency of any decision model in Cloudflare's Decision Index runs at 38.8 ms median.
Jev Models & Reimplementations#community#open-models#local
Community
Bespoke Labs' 9B decision model on Ollama, fine-tuned from Qwen3.5-9B: Apache-2.0, up to 64 typed questions per call, answers in under 100 ms on a MacBook Pro with an M5 Max, and 75.7% mean accuracy across 13 human-labelled datasets.
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.