Introducing Versatile 1.1

We're introducing Versatile 1.1, a zero-shot classification engine from Korollr. You give it a piece of content and a set of labels, and it tells you which one fits - no training, no fine-tuning, no examples. It works across text, image, audio, and video, and it answers in three typed shapes instead of free text: Choice (pick one of N options), Score (a graded position on a scale), and Noul (a yes/no probability).

Versatile is available today through Korollr chat and the Korollr API, under the same key as our Lambert models.

Starting October 6, it is also available as a standalone package: npm install @athenox/versatile runs the same engine locally, no API key, no data leaving your machine.

Here's what you can expect:

• Three speed tiers. low, high, and max - pick the tradeoff that fits your latency budget. Moving up a tier costs more time and buys more accuracy; we measure both so you don’t have to guess.
• Multimodal. Text, image, audio, and video through the same decision layer. On external image and audio benchmarks, Versatile lands within a point of published reference scores for models built specifically for that modality.
• Long context. We’ve measured Versatile scanning documents past ten million tokens for a small set of matching passages. The architecture has no fixed context ceiling.
• Small and self-hostable. 129-540 MB depending on the tier. It runs in-process - no external API call, no data leaving your own stack, nothing to audit but your own deployment.
• Typed answers you can act on directly. A Choice question returns the selected option plus a probability for every option you offered. A Score returns a position on your own scale - it can land between two levels, not just on one. A Noul returns a single probability. Nothing to parse out of generated text.

Benchmarks (BTZSC full test sets, 12,480 examples; CIFAR-10 and ESC-50 with external ground truth):

• low - 5.5 ms, 0.588 average accuracy
• high - 10.6 ms, 0.608 average accuracy
• max - 23.5-34.2 ms, 0.634 average accuracy
• CIFAR-10 (image) - 0.890, published reference 0.90
• ESC-50 (audio) - 0.818, published reference 0.82

Per-benchmark, Versatile High vs. published Jev numbers:

• agnews - Versatile High 0.769, Jev 0.910
• banking77 - Versatile High 0.612, Jev 0.870
• emotion - Versatile High 0.522, Jev 0.480

Jev leads the first two by a wide margin. On emotion, Versatile comes out ahead.

What it's for:

• Content moderation - text, image and audio through the same pipeline
• Ticket triage - urgency, category and sentiment in one call, without customer data leaving your own stack
• Fraud pre-screening - score every transaction before escalating the borderline ones
• Agent routing - decide which sub-agent or tool to call without paying full-LLM cost for a simple branch
• Sentiment analysis

A structural test, clearly labeled as such: we ran Versatile against the problem shape behind content moderation - finding a rare category inside a long stream of images. Using real photographs with external ground truth, it recovered 14 of 15 planted targets (93%) with a 0.2% false-positive rate across the surrounding images. This isn’t a moderation benchmark - telling a cat from a ship is a much easier problem than telling a violent scene from an action movie - but it shows the retrieval mechanics work: rare-class recall stays high even as the target gets sparser in the stream.

Versatile is free while in beta. Pricing may change - we’ll announce it ahead of time, and existing keys will keep working.