Frontier output, a fraction of the tokens. A yes/no decision doesn't need a frontier model. Neither does checking whether a document is relevant. @smartloop sends each part of a request only as far as it needs to go. 01 Split the request A request becomes small typed steps: collect, judge, extract, tool, fetch, reason. Each step gets only the context it needs. No step gets the whole conversation. 02 Spend tokens where they count Decisions run on the fast System 1 model, capped at 64 tokens. Tool resolution gets room to reason. Only content steps load a bigger capability model. 03 Filter before you prompt Retrieved context is judged for relevance before it enters a prompt. An unrelated document costs zero tokens. 04 Share weights, skip the bill One model host keeps weights loaded across every project. If a model is skipped, the stream says why. Here is what local first means. If your machine has the resources and the model, the step runs on your machine. If it doesn't, the step is delegated to a cloud model. Open and frontier models run as one system. On real hardware: Contract review on an M4 Mac, 24GB. Orchestration, hybrid search, model selection and eval on a five-year-old Linux machine running Qwen. Built in Rust on llama.cpp. Quantized open models on macOS, Linux and Windows. MCP tools, reusable skills, agents, OpenAI-compatible API. Studio 1.0.58 added Legal and Healthcare templates, including contract review, plus one-click MCP connections. It started as private document processing for European financial institutions. Embedding every document into a single model hit its limits, so I split the work into steps. That was two years ago, mostly nights and weekends. I'm looking for a few enterprise design partners in privacy-sensitive industries, starting with finance and law. Bring a real workflow and we'll build around it. #LocalAI smartloop.ai/docs

Frontier quality. Local First @smartloop smartloop.ai/?s=x

If you're using doc-to-chat wrappers and sending your personal and legal documents to a product you don't know, do yourself a favor: download @smartloop from the App Store, pick a #Qwen model from the list, add your documents, set up your MCP and skills, and optionally enable AI search. All of this costs $0 (unless you need AI search <=$5), and your data stays with you. #DocToChat