Are the reported benchmark numbers independently verified?

Treat them as author-run signals: JevBench rows are self-scored with the official harness and Decision Index work cites a pending validation PR, not a leaderboard result.

Which should I deploy, v1.1 or v1.2?

Pick v1.2 for distracting or adversarial inputs with roughly half the flip rate, or v1.1 for clean inputs where the last accuracy points and evaluated thinking matter.

What hardware do I need for self-hosting?

It needs an NVIDIA GPU Linux host with vLLM, 8.4 GB for BF16 weights or smaller GGUF files for llama.cpp, Ollama and LM Studio.

Does it contain Jev weights or need the TypeSafe API?

No. The digest states it is an independent alternative with no Jev weights and no TypeSafe affiliation, sharing only the API shape.

How do I start the Wald-4B server locally?

Download the v2.0 revision with hf download org2ai/Wald-4B --revision v2.0 --local-dir ./Wald-Q4B-v2, then run ./run.sh "$PWD". It serves a Jev-compatible POST /v1/systemone on port 8000 using uv and one NVIDIA GPU.

When does Wald-4B think before answering?

In the default Auto 0.7 mode, about one third of requests trigger a short native thought of at most 512 tokens when the top probability is below 0.7, then it reads the options again. One pass is deterministic, but answers after a thought are sampled and can vary slightly.

Can Wald-4B read images as well as text?

Yes. Images go through Qwen3.5-4B's vision tower unchanged, zero-shot, and the README reports about +2.8 points over its base on CV-Bench, BLINK, and RealWorldQA. You pass an images array of data URIs alongside the state and questions.

CommunityJev Agents & MCP

Wald 4B

Open-weight 4B decision model on Qwen3.5-4B-Base serving a Jev-compatible /v1/systemone API, returning per-option probabilities with adjustable thinking.

See highlights

Open github.com

Stars: 6
Language: Python
License: APACHE-2.0
Published by: Community
Author: org2AI
Last verified: Oct 2026

org2AI/wald-4b

Organization repository on GitHub

View on GitHub

Wald-Q4B: open-weight 4B decision model. Calibrated probability for every option, Jev-compatible /v1/systemone API, self-hosted. Weights on Hugging Face.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

license: apache-2.0 basemodel: Qwen/Qwen3.5-4B basemodelrelation: finetune libraryname: transformers pipelinetag: text-generation language: name: Wald-Q4B v2 (04701-c22, revision v2.0) results: task: type: text-classification name: Structured decisions (option probabilities) dataset: name: Decision Index 0.3, complete public suite (140,620 requests) type: decision-index revision: 62d2f51de34a2de64906345b6bc3e98e27ff55c7 metrics:

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

4B Qwen-based weights under Apache-2.0 with choice, noul and score questions.
Offers none, low, medium, high and multi-thought efforts plus calibration tables.
v1.2 halves adversarial flip rate while v1.1 keeps slightly higher clean accuracy.
Serves Jev-compatible /v1/systemone with GGUF builds for CPU and consumer GPUs.

Watch out

Independent open weights with no Jev weights and no TypeSafe affiliation; benchmark rows are author-run and await maintainer validation. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

agent-harness-with-jev-llm

Demo of the LangChain Jev harness pattern in LangGraph: a fast classifier routes events and gates a risky tool call, leaving open-ended reasoning to a chat LLM. Entry is POST /events; output lands in JSONL files.