org2ai/Wald-4B

Wald-Q4B v2

An open 4B decision model: send a state and typed options, get a calibrated probability for every option. Served through a Jev-compatible POST /v1/systemone API on your own GPU.

Why Wald-4B v2

• Probabilities, not text: a calibrated probability for every option, nothing to parse.
• Fast: median 42 ms one pass, 148 ms Auto 0.7 (serial, one RTX PRO 6000).
• Thinks only when unsure: about one third of requests trigger a short native thought.
• JevBench public: 210/231, against 200/231 for Jev 1.13.
• Decision Index 0.3: 54.18 on the complete public suite (our run).
• Reads images zero-shot: +2.8 points over its base on CV-Bench, BLINK, RealWorldQA.
• Small and open: 4B, Apache-2.0; GGUF from 2.7 GB in org2ai/Wald-4B-GGUF.

Scores

Benchmark	Wald-Q4B v2 · one pass	Wald-Q4B v2 · Auto 0.7	Jev 1.13 (hosted)
Decision Index 0.3, complete public suite	not run	54.18 (our full run)	57.96 (board public column)
JevBench public set (231)	204/231	210/231	200/231
JevBench-XL TEST (8,177)	65.20 %	65.78 %	67.85 %

Self-run, not leaderboard results; how each number was produced: evaluation/v2/.

Quick start

hf download org2ai/Wald-4B --revision v2.0 --local-dir ./Wald-Q4B-v2 && cd Wald-Q4B-v2
./run.sh "$PWD"          # POST /v1/systemone on :8000 (one NVIDIA GPU, uv)

curl -s localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "The customer wants to return a damaged kettle.",
  "questions": {"route": {"type": "choice", "instructions": "Choose the support queue.",
    "criteria": {"returns": "Returns and refunds", "delivery": "Delivery tracking", "other": "Other"}}}}'

curl -s localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Photo from the returns desk.", "images": ["data:image/jpeg;base64,..."],
  "questions": {"damaged": {"type": "noul", "instructions": "Is the item visibly damaged?"}}}'

API fields: docs/api.md. Server options, Docker and evaluation settings: RUNBOOK.md.

How it works

• One forward pass reads a probability for every option from the model's option-letter logits (Qwen chat template), calibrated with a frozen temperature table.
• Auto 0.7 (default): if the top probability is below 0.7, the model thinks once natively (at most 512 tokens) and reads the options again.
• More than 26 options: a one-pass knockout over all options.
• Images go through Qwen3.5-4B's vision tower, unchanged (zero-shot).

Details

• Training. Full-parameter decision training from Qwen3.5-4B (chat); checkpoint 04701-c22.
• Data. About 15% of training tokens come from the train splits of public benchmarks that the Decision Index also draws on, reformatted as decisions. The rest is broad decision data: general decision tasks — rules and policies, classification, entailment and fact checking, tables and documents, web and tool actions, judging (about 40%); math and step-by-step reasoning (about 20%); and hard decisions written and/or labelled by large frontier models, including robustness cases with perturbed or adversarial states (about 25%). Public datasets are used as train splits only. Contamination scan against the full Decision Index 0.3 public suite: 0 strict hits.
• Limits. Scores are self-run; confidence is not correctness, so set thresholds on your own data. Answers after a thought are sampled and can vary slightly; one pass is deterministic.
• Licence. Apache-2.0 for weights and code (NOTICE). Earlier versions: tags v1.2, v1.1, v1.0. Cite: CITATION.cff.