Krus 2 and a faster Ichnus
A deep dive into what makes Krus 2 cheaper and faster to run — and why that is the whole point.
Krus 2 shipped this week. The user-facing story is simple: Ichnus got noticeably faster. The interesting story is underneath, because Krus 2 is not just a better model — it is a cheaper one to run, and at Curos those are the same thing.
Why serving cost is a mission problem
Ichnus is free, which means its serving cost is not an operating detail; it is a constraint on the mission. Every request answered by a big model is a request we serve at a loss of margin. The difference between serving a request for a penny and serving it for half a cent is the difference between an assistant that can be free for everyone and one that can only be free for a while.
So when we improve a model, the metric we care about most is often not the benchmark score — it is the cost per useful answer.
What changed in Krus 2
Three things, all of them about spending less compute on the same work.
Tiered routing got smarter. Our routing layer (the one that sends easy questions to small models) now predicts request difficulty with a more conservative threshold and a re-check pass. Krus 2 handles more requests directly, and the ones it hands up to Mantus are genuinely hard — not just slightly long. The share of traffic answered by the largest models dropped again.
Attention got cheaper. We applied our quantized-attention work to the serving stack. Krus 2 runs attention at 4-bit precision with a quantized key-value cache. For the short contexts that dominate everyday traffic, this is where the latency wins are — Krus 2’s first token arrives about a third faster.
The model itself is leaner. Krus 2 is more capable than Krus 1 at the same size, which is a strange sentence to write, so let me be concrete: on our internal Curos-Helpfulness evaluation it scores higher while costing less to serve, because the architecture wastes less compute on irrelevant context.
What it means for you
If you use Ichnus, Krus 2 mostly means “it feels faster” — quicker replies, snappier drafting, less waiting on routine questions. If you use Ichnus on a train, it means the app does more locally before touching the network.
If you care about the mission, it means something else: the single largest cost of keeping Ichnus free went down, and the headroom came from engineering, not from a new funder. That is the direction we want to keep going.
All news →