The AI behind QuickCasa just got a big upgrade

The AI behind QuickCasa just got a big upgrade. QuickCasa runs on Qai, our own model layer, and every text task now routes through a new generation of three core models: qai-pro, qai-max, and qai-think. If you like knowing what is under the hood, here is the technical version.

All three speak the same OpenAI-compatible API, stream token by token over SSE, accept system prompts, and honour the usual sampling controls (temperature, top_p, stop sequences). Where they differ is architecture, context, and how they spend compute.

qai-pro: the balanced default

qai-pro is a sparse mixture-of-experts model. It holds a large pool of parameters but activates only a small slice of them per token, which is how it stays quick (first token in roughly 400ms) while still handling real reasoning, code, and tight structured output. It follows instructions precisely, including negative constraints, and reliably returns clean JSON against a schema.

The new generation also widens its context window substantially, so a long multi-turn conversation stays coherent instead of drifting once it runs past the old limit.

In QuickCasa: this is the model behind most live renter conversations and the qualification step. Faster first tokens mean snappier replies; better schema adherence means cleaner captured fields (budget, move-in date, pets) with fewer retries.

qai-max: long context and deep reasoning

qai-max is the heavy tier. It scales the same mixture-of-experts idea up to a much larger active-parameter budget and pairs it with a compressed-attention design that makes very long context practical: it handles up to a million tokens in a single request while using a fraction of the memory and compute the previous generation needed at that length.

That combination is what lets it hold an entire portfolio's worth of inventory, a renter's full history, and your brand guidelines in one prompt and still reason over all of it. First token lands in roughly 700ms, the trade you make for depth.

In QuickCasa: qai-max handles the long-context and high-stakes work: generating listing copy and emails that stay on-brand across many paragraphs, and answering complex questions that need the model to weigh a lot of context at once.

qai-think: explicit reasoning and tool use

qai-think is a reasoning model. Before it answers, it produces an internal chain of thought inside dedicated think tokens, works the problem step by step, and checks itself, so its final answer is the product of deliberation rather than a first guess. It is tuned for agentic loops: it plans, calls tools, reads the result, verifies, and retries when a step fails, holding the thread across a long tool chain.

It also carries a large context window and an unusually big output budget, so it can produce long, structured results without truncating.

In QuickCasa: qai-think is reserved for the tasks where a confident wrong answer is the real failure: matching a renter's exact criteria against dozens of units, computing pro-rated rent, or resolving a tour time across several constraints. It is slower on purpose, and worth it when correctness matters more than latency.

How QuickCasa routes between them

You never pick a model. QuickCasa routes each task to the cheapest, fastest tier that can still do the job well: qai-pro for the conversational default, qai-max when quality has to hold up unedited or the context is large, qai-think when the problem needs step-by-step reasoning. High-volume, low-stakes work like top-of-funnel triage and classification drops to a lighter tier again. The routing happens per task, invisibly, so you get the right trade-off on every call without configuring anything.

The net effect: lower latency on the common path, more headroom on the hard path, and better answers across the board.