One binary and one compatible endpoint: point the client at it and the code does not move while the model comes home. Engines arrive only when a model needs them, every feature has a path that runs on the processor, and a second machine becomes more capacity.

Cutting the key-value cache into blocks instead of reserving it whole lifts memory use from three or four tenths to above nine, and the same card then serves two to four times as many requests - which is the difference between a demo and a service.

Drag PDFs, word-processor files and tables into a workspace and you get an assistant that has read everything you own, with a citation under every answer - and the model, the vector store and the storage all live on the machine by default.

Written in plain C with no dependencies, it set the file format that made quantised models small enough for ordinary laptops, and it is the layer several familiar desktop tools quietly build on.

No discrete card, no account, no command line: a quantised open model runs on an ordinary processor, and a folder of your own paperwork becomes a question-answering collection that never leaves the machine.

Every line of the chat window, the model manager and the local server is published under a permissive licence, so the conversation can run with the cable pulled and nobody has to take the privacy claim on faith.

A runtime rather than a chat application: one command pulls and runs an open model, a local endpoint speaks the common request format, and any script or editor plugin you already have can point at it without touching the outside network.

A client of roughly five megabytes puts the models you already pay for into one window, keeps every conversation in the browser, runs on six platforms, and charges nothing itself because the bill goes straight to the provider.

Point it at a model server you already run or at any compatible API, and one browser window becomes a workspace with side-by-side models, a document knowledge base, role permissions and single sign-on, able to run with no outside network at all.

Free software you run yourself: one window for several model providers, agents assembled without code, a sandboxed interpreter and MCP connectors, with cost per team member visible instead of dissolving into a fixed subscription.

One subscription covering both the daily work and the code: a plan first, execution across connected tools afterwards, answers at about a thousand words a second, and data hosted inside the European Union by default.

Drop a million characters of material in and it splits the work itself, hands the pieces to a swarm of helpers, and returns something you can use. Reading that much was the older story; finishing the job is what this generation is about.

A model with five hundred and fifty-two billion parameters that reads your screenshots without being told what is in them, takes a million tokens in one pass, and bills cached input at a fiftieth of a yuan per million. The interesting part is not the size but where the cost went: reading and writing now run on separate sets of weights.

Cherry Studio is a cross-platform desktop client for AI conversation: ask several models at once and compare, keep your own material in a local knowledge base, and generate images or translations beside it.