In development · Q3 2026 target

PNEUMA · PROPER NAMES UNDERSTANDING MODEL

One model that understands every proper name.

A hybrid CNN-Transformer for sequence classification and token-level parsing — entity type, language, country, gender, and name structure, all in one forward pass. Designed to run on a single CPU core at production scale.

Read the Onomas-CNN X paper ↗ Use PNEUMA-DD today →

TASK SURFACE

Five tasks, one inference.

PNEUMA produces calibrated outputs for every task that operates on a name string. Each task is independently trained on MondoGraph supervision; the shared encoder amortizes computation.

ARCHITECTURE

CNN front-end, Transformer body, multi-head output.

The CNN front-end gives the speed of Onomas-CNN X; the Transformer body gives the accuracy of fine-tuned XLM-R. A shared encoder serves five task heads. The model is designed to fit in 200MB and run inference on a single CPU core at 2,500+ names per second.

1
INPUT
UTF-8 name string, script-tagged
2
CNN STEM
Parallel depthwise-separable convolution branches (1×3, 1×5, 1×7), 64ch
3
TRANSFORMER BODY
4 layers, 8 heads, d=256 — attention over CNN-extracted tokens
4
FIVE HEADS
Entity, language, country, gender, parsing — calibrated outputs
5
CALIBRATION
Temperature scaling on held-out per locale (ECE < 0.02 target)

Legacy & lineage

JAN 2026 arXiv:2601.11090

Efficient Multilingual Name Type Classification Using Convolutional Networks

Davor Lauc · the CNN-only precursor to PNEUMA

92.1% accuracy 104 languages 2,813 names/sec/CPU 46× faster than XLM-R

Demonstrated that specialized CNN architectures remain competitive with fine-tuned large LMs on focused NLP tasks when sufficient supervision is available. PNEUMA extends this with a Transformer body for tasks where context matters (parsing, gender), and a multi-task head for joint training.

2021 · EACL BSNLP

A Pre-trained Transformer for Croatian, Bosnian, Serbian and Montenegrin

Ljubešić & Lauc · 8B-token Slavic transformer · informs the Transformer body

USE IT TODAY

PNEUMA-DD ships every PNEUMA task in production.

The data-driven variant uses MondoGraph token statistics directly to serve every classification and parsing task. Same API, same input schema, ready today. When PNEUMA reaches Q3 2026 GA, it drops in behind the same endpoint with no client changes.

PNEUMA-DD →
# One call, all five tasks. curl -X POST https://api.mondonomo.ai/v1/pneuma \ -H "Authorization: Bearer $TOKEN" \ -d '{"name": "Eugen Schochenmaier"}' # → { "entity_type": "person", "language": "de", "country": "DE", "gender": "masculine", "parts": [ {"slot":"given", "value":"Eugen"}, {"slot":"surname", "value":"Schochenmaier"} ] }