Jev directory
CommunityJev Models & Reimplementations
Soro-Kev-4B
A pointer-network decision model from Keyush Nisar that reads one agent state document plus typed questions and returns calibrated tool-routing probabilities in a single forward pass, reporting 89.00% zero-shot tool routing.
- Reading time
- 3 min
- Published by
- Community
- Author
- nisaralll
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- MCP
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- LoRA adapter (r=16, alpha=32) plus a pointer head on a frozen Qwen/Qwen3-4B, served in NF4 4-bit at about 4,048 MB VRAM.
- Reports 89.00% zero-shot tool routing on its Universal Tool Harness and 94.50% on external human holdouts.
- Median prefill latency is about 413 ms on an RTX 4050 Laptop GPU or Kaggle T4; fp16 serving needs about 8,200 MB.
- Learned temperature 1.62 and a split-conformal quantile of 0.8066 at 90% coverage drive EXECUTE, ASK and abstain outputs.
- Handles choice and noul question types over states up to 32,768 tokens, though training states were capped at 4,096.
Watch out
Card reviewed 2026-10-17; the 89.00% and 94.50% routing figures and the 75.76% BFCL irrelevance rejection are self-reported on the card with no third-party reproduction. Training states were capped at 4,096 tokens (8,192 validated) even though the card advertises 32,768-token states, and Apache-2.0 covers the adapter and head while the Qwen3-4B backbone keeps its own licence.
Questions about this entry
Short answers drawn from the linked project or page.
A decision model for agent tool routing, multi-protocol schema normalization and execution boundary control. It reads one document (the state: prompt, conversation history and registered tool schemas) plus typed questions, and returns a calibrated probability distribution over the supplied options in one forward pass without generating text.
It follows the same shape: you supply a state and typed questions, and it answers with probabilities over your options instead of text. Each question branches from a shared cached state representation, so questions cannot cross-contaminate one another and the state prefix is computed once.
It is a LoRA adapter (rank 16, alpha 32) and a pointer head over Qwen/Qwen3-4B. The card documents two serving modes: NF4 4-bit at about 4,048 MB VRAM or fp16 at about 8,200 MB, with evaluation performed in fp32.
The card reports a median prefill latency of about 413 ms on an RTX 4050 Laptop GPU or a Kaggle T4 GPU, and targets production agents with large MCP catalogs of 50 or more tools; it does not publish an install or serving command.
The card claims 89.00% zero-shot tool routing on the Universal Tool Harness, 94.50% on external human holdouts and 75.76% calibrated abstention on BFCL irrelevance slices. No third-party reproduction is linked, so treat these as the author's own evaluation.
Keyush Nisar built Soro-Kev 1.0 (release v0.2.0, dated 2026-10-17) on the pointer-network methodology of jaredpalmer/kev-0.8b, directly addressing the original Kev cards' warning that tool-call routing accuracy was below chance at 0.133.
Beyond tool routing it maps argument keys to schema properties, checks whether required arguments are missing and classifies whether an action mutates external state. Its conformal prediction set controls execution: a singleton set triggers EXECUTE, two to eight options trigger ASK, and abstention halts the call.
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 local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.
Jev Models & Reimplementations#community#rust#open-models
Community
ollaya-dev/ollayaRust
ollaya-dev/ollayaRustThe first v1 checkpoint of the Strands decider line from the Strands Agents team: a LoRA adapter and readout head on Qwen3.5-2B-Base that answers typed decisions with calibrated confidence, reporting 0.7792 JevBench public accuracy (180/231).
Jev Models & Reimplementations#community#open-models#local
Community
A Strands decider v1 checkpoint: a LoRA adapter and readout head on google/gemma-4-E4B-it that answers typed decisions with calibrated confidence; its model-index reports 0.8095 JevBench public accuracy (187/231) and mixed code-bug evaluation scores.
Jev Models & Reimplementations#community#open-models#local
Community
A Strands decider v1 checkpoint: a LoRA adapter and readout head on google/gemma-4-E2B-it that answers typed decisions with calibrated confidence; its model-index reports 0.7792 JevBench public accuracy (180/231) and code-bug evals from 0.281 to 0.978.
Jev Models & Reimplementations#community#open-models#local
Community
A 2B decision model from Strands Agents for agentic workflows: a LoRA adapter on Qwen/Qwen3.5-2B-Base with a readout head that answers noul, choice and score questions with calibrated confidence, reporting 0.7619 on JevBench at 4096 context.
Jev Models & Reimplementations#community#open-models#local
Community
Jared Palmer's 27B Kev decision model: a full-weight fine-tune of Qwen3.8-27B averaged with v1, plus a pointer head that returns calibrated probabilities for typed choice, score and noul questions in one forward pass. Card-reported transfer-v4 accuracy is 88.87%.
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.