Krus
Light and quick. Built for everyday questions, fast replies, and running at scale without the wait. Krus handles the small stuff so the bigger models stay available for the hard stuff.
Best for
- Quick lookups and factual questions
- Drafting and editing
- Simple chat and customer-facing flows
- High-volume, low-latency use
Why this tier exists
Most questions people ask an assistant do not need a frontier model — they need an answer in under a second. Krus exists so that the fast, common case stays fast, and so serving it doesn't compete for the compute that Corus needs for hard problems.
Being the cheap, fast tier is not an excuse to cut corners on judgment. Krus passes the same safety and honesty suite as Corus before release — it just runs a smaller model, tuned hard for latency, not a less-tested one.
Benchmarks
Scores on public benchmarks and two internal evaluations. Figures are illustrative — see the note below.
| Evaluation | Krus | Krus | Mantus | Corus |
|---|---|---|---|---|
| MMLU | 78.9 | 78.9 | 85.6 | 88.2 |
| GPQA (diamond) | 58.2 | 58.2 | 67.8 | 76.4 |
| MATH (competition) | 62.4 | 62.4 | 71.0 | 79.1 |
| HumanEval | 76.5 | 76.5 | 84.1 | 89.3 |
| OversightQA (internal) (ours) | 71.8 | 71.8 | 79.4 | 84.6 |
| Curos-Helpfulness (internal) (ours) | 80.2 | 80.2 | 86.9 | 88.7 |
Benchmark scores are illustrative and invented for this concept project. Curos is not a real organization and these are not measurements of a real model.
Model card
The sections below mirror the model-card format we publish for every release. For Krus, the release report is published alongside the evaluation suite it passed.
Overview
- Version: Krus 2, released January 2026
- Context window: 128k tokens
- Availability: available to all Ichnus users, free of charge
- Safety classification: consumer general-purpose assistant, not for high-stakes uses without review
Training and data
Krus was trained on a curated corpus, with a strong emphasis on removing low-quality and duplicated data. We publish an outline of our data practices in the model documentation and the details behind each release in its evaluation report.
Known limitations
- May confidently answer outside its knowledge — we are working on honesty, not pretending it is solved
- Not a substitute for professional advice in medicine, law, or finance
- Can produce plausible but wrong code; always review generated code
Safety evaluations
Before release, Krus passed our full evaluation suite — capability, safety, honesty, and red-teaming — as reviewed by the Safety & Evaluations Committee. The public report lists what we tested, what we found, and what we did not test. See our approach to evaluations for how the suite works.
Where to use it
- Lowest latency in the Ichnus family
- Best for chat, quick lookups, and drafting
- Available to every Ichnus user by default
- Optimized for efficiency in serving