Skip to content

Choose models with pasture onboard

pasture onboard finds models that fit this machine and the work you will do with them, and downloads the ones you choose. It reads the machine's CPU count, architecture, RAM, available memory, and the disk that holds Ollama's models. On Linux it also checks memory limits under the usual cgroup mounts. It detects Apple silicon unified memory, also when pasture runs under Rosetta, and queries NVIDIA or AMD video memory when those measurements are available. GPU detection does not verify that Ollama supports the device or that its driver works.

pasture onboard
pasture onboard --describe "Help me write software and search documents"
pasture onboard --use-cases coding,embeddings
pasture onboard --offline --json --use-cases coding,vision

pasture install offers to run onboarding after it asks about pairing, on a terminal and when the library has no models yet.

Answer the questions

On a terminal, onboarding shows a checklist of use cases: Coding, Reasoning, Chat and writing, Reading images, Search and RAG, and Creating images. Move with the arrow keys, check a use case with Space, and confirm with Return. To say what you want in your own words instead, check the last choice, "Describe it in my own words", and type a sentence such as "review pull requests and search our docs". Esc on that sentence goes back to the checklist, and Esc or Control-C on the checklist stops onboarding.

--use-cases names the use cases and --describe gives the sentence, so onboarding skips the checklist. --workloads, its name before 0.2.0, still works. The use case names are coding, reasoning, general (chat and writing), vision (reading images), embeddings (search and retrieval), and image-generation (creating images). The JSON report keeps them under workloads, as it always has.

Away from a terminal, as in a script, onboarding asks one question and reads one line: the use cases, by name or number and separated by commas, or a sentence about the work. --json asks nothing and prints the plan as JSON.

Where the recommendations come from

pasture asks Cinchapi to recommend models. It sends the sentence, the checked use cases, the machine's capacity (the operating system, architecture, CPU count, memory, free disk, and each GPU's name and free video memory), and the names of the models the library already has. It sends no host name, user name, file, chat, or credential. Cinchapi answers with the best models for each use case and the size of each. pasture checks each one against its own memory and disk limits, and passes over one that does not fit or that needs a license you accept first. The plan then says "Recommended by Cinchapi", with each model's name and why it suits the work.

When Cinchapi does not answer within five seconds, refuses, answers in a form pasture cannot read, or recommends nothing that fits, pasture plans from its own catalog instead and says "Recommended by pasture's own catalog". --offline plans from the catalog without asking Cinchapi. For the catalog, keywords in the sentence add use cases: a keyword matches only at the start of a word, so "storage" does not count as "rag", and a sentence that matches no keyword plans for chat and writing.

The models_url key in ~/.config/pasture/config.json points onboarding at another Cinchapi server, such as a staging one.

Read the plan

The plan lists each model with its use cases, the disk it takes, and the memory it needs while it runs. A plan from Cinchapi takes, for each use case in turn, the best-ranked model that fits the disk the models before it left. A plan from the catalog reserves space for smaller models across the chosen use cases before spending spare disk on larger ones. The downloads together must fit above the free-space reserve and within the unused part of the budget. The plan does not count on moving installed models away.

On a terminal, onboarding then asks what next:

  • Download these models downloads each one through pasture, which makes room for it first, with a progress bar, and suggests a pasture run command when they are done. pasture must be running.
  • Change my answers goes back to the checklist.
  • Done for now downloads nothing.

Away from a terminal, the plan ends with a pasture add command for each model, and onboarding downloads nothing. Onboarding never changes pasture's settings.

Without a GPU or Apple silicon, a model that reasons before it answers, such as qwen3.5, can take minutes for its first answer, and can fill a 4096-token context before it answers at all. For such a model, the plan gives a pasture run command with --think=false, which turns reasoning off. In the JSON, "thinks": true marks such a model, and "accelerated" says whether the plan counts on a GPU or Apple silicon.

Memory estimates assume one active model, a 4096-token text context, and room for the OS and runtime. CPU-only plans favor smaller models. Discrete GPU memory is never added to RAM or summed across GPUs. A GPU with less free memory than the CPU-only plan allows, such as an integrated GPU, leaves the CPU-only plan in place, because Ollama runs a model that the GPU cannot hold on the CPU. Larger contexts, image inputs, concurrent requests, drivers, and model updates can change actual requirements. Unknown available memory produces no fit recommendation.

Browse every model

pasture browse lists models from Cinchapi's directory, most popular first, with what each is for, the disk it takes, and how it fits this machine:

pasture browse
pasture browse coder
pasture browse --use-cases coding,embeddings
pasture browse --all --json

For each model, pasture picks one file. It prefers the Q4_K_M build, which keeps nearly all of a model's quality at about a third of the full size and answers about three times as fast. For a model of under 4B parameters, or one for search and RAG, it prefers Q8_0, since quantization costs a small model more. It takes the first file in that order of preference that fits, and the full-size F16 or BF16 build only when a model has no other. It never offers a build of fewer than about 3.4 bits per weight, such as Q2_K, whose answers suffer too much. Then it labels the model:

  • runs well: it needs at most four fifths of the memory one model may use.
  • tight: it needs nearly all of that memory, so longer contexts may not fit.
  • needs room: it fits in memory, but the disk has less room than it takes.
  • too big: no file of it fits in memory.
  • no download: pasture cannot download any file of it, because every file asks you to accept a license first, none is in a format that Ollama runs, or every one keeps too few bits per weight.

By default, the list holds only models that fit in memory. --all lists every model, --use-cases keeps the models for any of those use cases, and words after the command search names and descriptions. pasture asks only for models with a file that Ollama can run, and sends Cinchapi only the search, the use cases, and the memory one model may use.

On a terminal, the list fills the window. Move with the arrow keys or Page Up and Page Down, press / to search, a to switch between models that fit and all models, and Return to pick one. pasture then offers to download it, go back to the list, or stop. The list loads more models as you scroll. Away from a terminal, pasture browse prints the first 50 models as a table with the name that pasture add downloads, or as JSON with --json.

When Cinchapi's directory does not answer, and with --offline, the list holds pasture's own catalog, and says so when the directory failed. Onboarding offers the list too, as Browse all models, after it shows its plan.

The bundled catalog

The bundled catalog uses explicit size tags from Qwen 2.5 Coder, Qwen 3.5, Nomic Embed Text, and FLUX.2 Klein, checked on October 30, 2026. Download estimates include headroom over the published sizes; runtime memory is a conservative estimate. Tags can change. The normal pull path checks the current registry size and enforces pasture's budget when you download. Image generation is experimental and the bundled recommendation requires macOS on Apple silicon. Other platforms report that limitation.