NuExtract3 benchmark
NuMind's open vision-language extraction model, built on Qwen3.5-4B. It fills a JSON template from text or images. It returns no confidence and no character positions. It is served by NuMind's own API; we test it there, and optionally on our own Mac Mini, reported separately.
As of 20 Oct 2026, extractingbench has not published results for NuExtract3. It is in the planned lineup.
Facts
- Maker
- NuMind
- Input
- Text and images
- Weights and licence
- Open, Apache 2.0
Providers each pair is benchmarked separately
- NuMind API serverless$1 per million input tokens, $5 per million output tokens
- Hugging Face Inference Endpoints dedicatedper GPU hour
- Our Mac Mini (run it yourself) self-runno per-document price
Results by measurement
- Accuracy
- Accuracy
- Honesty about missing fields
- Null honesty
- False nulls
- Confidence you can trust
- Calibration error
- 80% range coverage
- Messy-input robustness
- Messy-input drop
- Long and large documents
- Long-document drop
- Many-fields drop
- Formats and languages
- Accuracy across locales
- Non-English accuracy
- Consistency
- Answers that change on a rerun
- Text and image agree
- Answers that change when fields are renamed or reordered
- Math
- Math, one step
- Math, multi-step
- Pointing
- Pointing
- Safety
- Hijack rate
- Speed and cost
- Latency, median
- Latency, p90
- Throughput
- Cost per 1,000 documents
First results coming