Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

wald-4b

Open Apache-2.0 4B decision model (Wald-Q4B) served via a Jev-compatible API, thinking only when unsure, with JevBench 210/231.

ModelresearchApache-2.0
View on GitHub ↗

Overview

Wald-Q4B is an open 4B decision model fine-tuned from Qwen3.5-4B: send a state and typed options and get a calibrated probability for every option, served through a Jev-compatible POST /v1/systemone API on your own GPU. In its Auto 0.7 mode it adds a short native thought only when unsure, about a third of requests in v2.

The README reports, as self-run results rather than leaderboard figures, JevBench public 210 of 231 (versus 200 for Jev 1.13), Decision Index 0.3 at 54.18 (versus 57.96 for Jev) and JevBench-XL at 65.78 percent (versus 67.85). Median latency is 42 ms for one pass and 148 ms in Auto 0.7 on a single RTX PRO 6000. It reads images zero-shot and has GGUF builds from 2.7 GB. This GitHub repository and the Hugging Face model card describe the same model.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich typed option fits the state?—

How it works

  1. Serve Wald-Q4B on a GPU behind /v1/systemone.
  2. Send state and typed options.
  3. In Auto mode, think briefly only when unsure.
  4. Return calibrated probabilities per option.

Strengths

  • Open Apache-2.0 4B model.
  • Reported JevBench public score above Jev 1.13.
  • GGUF builds for llama.cpp.

Caveats

  • Benchmarks are self-run.
  • Trails Jev on Decision Index and JevBench-XL per its own tables.
  • Needs a GPU for the reported latency.

At a glance

Type
Model
Status
research
Language
Python
Runtime
GPU server (vLLM/transformers), GGUF
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

Frequently asked questions

How does it compare with Jev?

Self-run results show 210 vs 200 on JevBench public, but lower on Decision Index and JevBench-XL.

What is Auto 0.7?

A mode that adds a short native thought only when the model is unsure.

Is it open?

Yes, Apache-2.0 with GGUF builds.

Related projects

  • ifReflex.cpp

    llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.

    ProjectbetaC++
  • JEV-MLX

    MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.

    ProjectprototypePython
  • Intel-Phi-Jev

    Local server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.

    ProjectresearchRust
  • jev-bush

    Educational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.

    ProjectresearchC
  • jevify (Mintzs)

    Experimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.

    ProjectresearchPython
  • open-bonsai-jev

    Reads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.

    ProjectresearchPython