← All troubleshooting topics

Models editor

The Models tab is a UI over the Ollama HTTP API. Most issues here are validation in the editor or something on the Ollama side: an unreachable server, a stalled download, or an error Ollama returns.

“Save as new model” rejects my Modelfile

Saving a derived model fails with a validation error.

The editor refuses Modelfiles that could smuggle extra directives.

• The Base model (FROM) tag must only contain letters, digits, ., _, /, or -, plus one : before the tag (so llama3.1:8b and a Hugging Face pull such as hf.co/user/repo:Q4_K_M are fine). No spaces, no quotes, and no empty path segments (a//b, or a leading or trailing /).
• The System prompt and Prompt template cannot contain """. If you need a triple-quote in your system prompt, rephrase.
• Save stays disabled until both New model tag and Base model (FROM) are filled in. If Ollama itself rejects the model - say, the base model isn’t installed - its error appears next to the button.
• Top-K, Context, Max tokens and GPU layers take whole numbers of 0 or more, and Seed a whole number. Anything else snaps back to the previous value.
• “Couldn’t read the Modelfile: …” - Ollama 0.5.5 and later want the Modelfile split into separate fields, and Loach couldn’t split this one. The usual cause is an unsupported directive: a System prompt or Prompt template that ends in a " runs into the closing """, and the leftover quote reads as a directive of its own. Add a space or a full stop after that last quote.
• Save as new model writes under the tag you type. If a model with that tag already exists - or the tag is the base model’s own - Loach asks before replacing it (Replace); pick a tag that isn’t taken if you want to keep the old one. Duplicate asks the same way before it overwrites a tag.

A model pull is stuck

The progress chip is hanging at the same percentage.

1. Click the ✕ on the progress chip and start the pull again. Ollama keeps the layers it already fetched, so the new pull picks up from there.
2. Confirm Ollama is still reachable - when it isn’t, the model picker in the chat header shows an amber warning next to Ollama, “Not running” and a Start Ollama button, and Settings → Providers → Test connection checks the base URL.
3. Check disk space on the drive Ollama uses for its model cache.

Can’t delete a model

Delete fails with a Couldn’t delete model toast.

The toast quotes Ollama’s own error - Loach never blocks a delete itself. Check that Ollama is still reachable (the chat-header model picker shows an amber warning and “Not running” when it isn’t) and that the tag matches what ollama list shows, then retry from the Models tab.

My per-model Thinking preference doesn’t apply

You set Allow thinking step in the Models editor, but a chat with that model still thinks (or doesn’t) the old way.

That setting is the default for chats with the model, saved on this device. A chat whose own Thinking toggle you flipped in its parameters panel keeps that choice - reset the chat’s parameters to fall back to the model’s default. The preference isn’t part of a backup either, so after Import data on another machine, set it again. Settings → Features → Thinking sets the default for models without a preference of their own.