Small models for exactly one task. As good as large ones on the main measures, many times faster.

We train our own models for decisions every AI application makes all the time: what does this message want? Is there anything personal in this prompt? They run on our hardware in Austria. This page shows what they do, how we measured them and where they still fail.

ALLSmartIntent

For chatbots and agents

ALLSmartIntent tells within milliseconds what a chat message wants and what a good reply needs. One call instead of many single questions to a large model, in 14 languages and for any bot.

≈ 15 ms per message, ≈ 29 ms on a CPU
message type 91.6% correct, on par with Clef 27B
18 answers in one call

Learn more →

ALLSmartArt4ALLSmartArt4iALLSmartArt9-10

For privacy (GDPR)

Before a prompt reaches an external model, three of our own models check it in the gateway's privacy layer: names and addresses, people without a name, special categories.

94.4% instead of 86.6% of personal values found (vs GLiNER)
15 ms instead of 75 ms for people without a name, without outliers
97.4% of sensitive texts found (Art. 9 and 10)

Learn more →

Why we train our own models

Large models are generalists

They can do almost anything but are too large for one narrowly defined check: every request through a model with 9 to 35 billion parameters costs hundreds of milliseconds to several seconds and a lot of GPU memory.

Specialised is faster

A small model that learns exactly one task gets it about as right as a large one on the main measures and is 5 to several hundred times faster, depending on the task. That lets us check every prompt without the check costing noticeable time.

The check stays in-house

A privacy check that first sends the text to someone else's model defeats its purpose. Our models run on our own hardware in Austria before anything leaves the house.

Used ourselves, shared openly

We use the models in our own applications and run them ourselves. Public versions are on Hugging Face.

How we train and measure

Synthetic and public data only

Training texts are invented, written by language models and by ourselves; ALLSmartIntent also learns from public datasets. No customer prompt goes into training.

Cross-checked

The labels of the training texts are cross-checked automatically. What stands out, we look at.

Measured blind

Test sets are written by separate agents that know neither the model nor the training data. Where we learned from errors, a new blind set is added.

Data stays in-house

We do not publish training data or test sets. We do publish method and numbers.

Open weights

Hugging Face · Apache 2.0

All four models have been published on Hugging Face under Apache 2.0. ALLSmartIntent is the same model there as the one we measure and use here. The three privacy models differ for licensing reasons: not all data we trained the versions here on may go into openly distributed weights. The public versions are therefore separate trainings using only sources that permit it, not the same weights as here.

ALLSmartIntent ↗ALLSmartArt4 ↗ALLSmartArt4i ↗ALLSmartArt9-10 ↗ ALLSmart on Hugging Face ↗

How close the public versions are to the ones at ALLSmartLLM

ALLSmartArt9-10 ↗

blind set 5: 151 sensitive, 149 harmless texts

	sensitive texts found	false alarms	precision
at ALLSmartLLM	147/151	7/149	95.5 %
on Hugging Face	144/151	14/149	91.1 %

ALLSmartArt4i ↗

blind set 2: 174 passages, 112 harmless texts; both versions without word filter

	passages found	harmless texts changed	precision
at ALLSmartLLM	86.8 %	5/112	95.7 %
on Hugging Face	86.8 %	5/112	92.0 %

ALLSmartArt4 ↗

blind set: 120 texts, 695 values

	values found	people found	precision
at ALLSmartLLM	94.4 %	97.4 %	94.4 %
on Hugging Face	93.8 %	96.1 %	92.1 %

All models run on our own hardware in Austria; the three of the privacy layer together in about 10 GB of GPU memory.