Maxim 2.0: The Brain Is Rented, the Body Is Owned
Alflora Research · 01 Aug 2026
Tl;dr
- Maxim 2.0 is a full AI chat app — accounts, conversation history, streaming — served from our own infrastructure at maxim.alflora.app.
- The inference is rented through OpenRouter and currently routes to Kimi K3; everything else is owned.
- The design is a direct answer to Maxim 1.0's autopsy: compute is a budget, so rent only the part that's expensive.
- Models are evaluated continuously and swapped when testing — not changelogs — says so.
The inheritance
Maxim 1.0's autopsy ended with a clear finding: compute is a budget before it is an architecture, and a bedroom rack cannot afford to train or serve a serious brain. But the autopsy also said something more useful — borrowed power is a bridge, owned power is the destination. Maxim 2.0 is that sentence turned into a product.
The 2026 revival makes a clean cut straight down the middle of the system. The expensive, power-hungry part — the inference itself — is rented per request through OpenRouter, currently pointing at Kimi K3. Everything that makes it a product instead of a demo is owned: the accounts, the conversation history, the streaming, the serving, the uptime responsibility.
What "the body is owned" actually means
Open maxim.alflora.app and everything you touch until the moment the model thinks is ours. Your login lives in our database. Your conversation history is stored on our disks, under our backup policy, deletable by you and readable by no one else. The UI streams tokens over infrastructure we restart at 3am when it misbehaves. If OpenRouter vanished tomorrow, the product would lose its voice — but not a single user, message, or byte of history. Those never leave the rack.
That is the whole point of the cut. A rented brain is a commodity: swappable, priced per million tokens, improvable without notice. An owned body is the actual asset: the users, the data, the trust.
Why the brain is safe to rent
Two reasons. First, evaluation is a process here, not a purchase order — providers and models are benchmarked continuously through OpenRouter and the routing changes when the answers do. Kimi K3 holds the seat because testing put it there, and it will keep the seat only while testing keeps it there. Don't believe the changelog; test it yourself.
Second, the privacy exposure of the rented hop is understood and bounded: the inference request is the only thing that crosses the bridge, and it crosses over TLS to a provider chosen with that exposure in mind. The crown jewels — who you are, everything you've said before — stay home.
The road back to owned
This is not the end state, and pretending otherwise would be dishonest. The destination from the 1.0 days hasn't moved: a brain we trained, served on hardware we own, answering under this roof. Maxim 2.0 is the bridge — it builds the product, the user base, and the evaluation muscle while compute economics catch up. When owned inference stops being a wall and starts being a purchase, the body will already be here waiting for its brain.
Borrowed power is a bridge. Owned power is the destination. Maxim 2.0 knows exactly which one it's standing on.