Jev directory
CommunityJev Models & Reimplementations
Nimble on Ollama
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.
- Reading time
- 2 min
- Published by
- Community
- Author
- Bespoke Labs
- Last verified
- Oct 2026
- Platforms
- CLI
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 9B decision model from Bespoke Labs, fine-tuned from Qwen3.5-9B; 44.2K downloads and 9.3-9.5 GB checkpoints with a 256K context window.
- Reads the prompt once per question and scores the answer tokens directly, with no reasoning step, which is what makes it fast.
- Up to 64 questions per request across three question types: choice, noul, and score, with 2 to 26 allowed answers each.
- Trained on contrastive pairs whose two examples differ by one fact, and that fact flips the answer.
- Apache 2.0 licence; 75.7% mean accuracy across Bespoke Labs' 13 public datasets (3,880 decisions) against Jev 1.13's 76.0%.
Watch out
Apache-2.0 weights. The headline accuracy is Bespoke Labs' own run, and the published held-out set is only 324 examples (162 contrastive pairs), where Nimble scored 292 of 324 (90.12%) against Jev 1.13.0's 302 of 324 (93.21%). The page shows no prompt-injection or multilingual testing.
Questions about this entry
Short answers drawn from the linked project or page.
A 9B decision model from Bespoke Labs, fine-tuned from Qwen3.5-9B. You send text and a set of questions; it picks an answer for every question and returns the probability of each allowed answer.
It reads the prompt once per question and scores the answer tokens directly, with no reasoning step. The page reports answers in under 100 ms on a MacBook Pro with an M5 Max.
Apache 2.0, per the Ollama model page. The default checkpoint is a 9.3-9.5 GB download with a 256K context window, and the library lists eight model entries in total.
Across Bespoke Labs' 13 public datasets with human labels, 3,880 decisions, Nimble 9B scored 75.7% mean accuracy, Tev1 4B 73.3%, Tev1 0.8B 63.5%, and Jev 1.13 76.0%. Nimble and Tev1 ran on Ollama; the Jev result came from TypeSafe's API.
Contrastive pairs in which the two examples differ by one fact and that fact flips the answer. Bespoke Labs held back 324 examples (162 contrastive pairs) from training to test on, and Nimble matched the reference on 292 of them.
The page shows a tool-call moderation example: render a call such as a shell command as state, ask a noul question about whether it could cause harm, and read the probability back before letting the call run.
The evaluation numbers are Bespoke Labs' own, the held-out set is small, and the page does not show prompt-injection, multilingual, or calibration testing. Confidence describes how concentrated the probabilities are, not the chance the answer is right.
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
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
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
Together AI's experimental decision family on Ollama, fine-tuned from Qwen3.5 at 4B and 0.8B: 93K downloads, up to 64 questions per call, an 88.0% development accuracy on 1,000 decisions, and MIT-licensed dataset builders and training scripts.
Jev Models & Reimplementations#community#open-models#local
Community
The typed-decisions checkpoint of the Laya family on ModelScope: a 421M ModernBERT-large fine-tune reporting 0.766 accuracy over 2,000 decisions, above TypeSafe Jev 1.13.0's published 0.727 and the 0.735 teacher ceiling, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
A 70.8M-parameter DeBERTa-v3-xsmall cross-encoder that answers typed decision questions in one batched pass with zero generated tokens, shipped with a calibration.json that the engine requires.
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
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.