org2ai/Wald-4B

Wald-Q4B v2.1

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 v2.1

• Probabilities, not text: a calibrated probability for every option, nothing to parse.
• Decision Index 0.3: 60.05 on the complete public suite (our run; v2.0: 54.18).
• Thinks only when unsure: Auto 0.7 thinks natively on about half of the requests.
• Fast: median 158 ms (serial, one request in flight, warm, processing time, one RTX PRO 6000, 200 DI 0.3 requests).
• Web actions: Mind2Web element selection 92.3 % (test splits, 6 candidates; v2.0: 84.6 %).
• Reads images and video zero-shot through Qwen3.5-4B's unchanged vision tower. Video (wald-serve 0.2.2): 87.4 % on VisualDecisionBench video (7,143 questions, our run), 2.5 s for a 15-second clip with 24 questions (serial, one RTX 5090).
• Small and open: 4B, Apache-2.0.

Scores

Benchmark	Wald-Q4B v2.1	Jev 1.13 (hosted)
Decision Index 0.3, complete public suite	60.05 (our full run; raw 70.12, breadth 58.63)	57.96 (board public column)
JevBench public set (231)	210*	200/231
JevBench-XL TEST (8,177)	67.75 %	67.85 %
Mind2Web element selection (test task / website / domain)	94.9 / 91.5 / 91.7 %	88.3 / 86.9 / 86.6 %
Image decisions, 11 public Decision Index Vision benchmarks (our conversions)	67.23	no image input
JevAdvBench flip rate, lower is better	7.1 %	6.1 %

* Measured on Wald-4B v2.0.

Self-run, not leaderboard results; how each number was produced: evaluation/v2.1/. DI 0.3 run files: org2ai/Wald-Q4B-decision-index-results.

Quick start

hf download org2ai/Wald-4B --revision v2.1 --local-dir ./Wald-Q4B-v2.1 && cd Wald-Q4B-v2.1
./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?"}}}'

curl -s localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Aisle camera, last 15 seconds.", "videos": ["data:video/mp4;base64,..."],
  "questions": {"conceal": {"type": "noul", "instructions": "Does the shopper put an item into a bag or pocket?"}}}'

Video needs the server from main (wald-serve 0.2.2; the weights are the same as tag v2.1): one video per request, up to 32 MB and 300 s. By default the server sends 32 frames to vLLM as one native video ("video_route": "native32"). "video_route": "frames16" sends 16 labelled frames as images instead, and is used automatically when the request also has images. Details: docs/api.md, numbers: evaluation/v2.1/video.json.

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). The server decodes a video once per request; by default 32 frames go to the tower as one Qwen3.5 video, with timestamps.

Details

• Training. One full-parameter pass of decision training from Qwen3.5-4B (chat), with a light KL anchor to the base model; the released weights are 0.85 × the trained model + 0.15 × the base (checkpoint 056A0-c31). Thoughts are the base model's own; no thought text was trained.
• Data. About 23% 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, judging (about 35%); hard decisions written and/or labelled by large frontier models, including robustness cases with perturbed states (about 15%); task-specific decision sets (about 11%); web and GUI actions (about 8%); and math and step-by-step reasoning (about 8%). 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. Thoughts are longer than v2.0's.
• Licence. Apache-2.0 for weights and code (NOTICE). Earlier versions: tags v2.0, v1.2, v1.1, v1.0. GGUF of v2.1: org2ai/Wald-4B-GGUF. The larger model in the same family is org2ai/Wald-26B-A4B. Cite: CITATION.cff.