← All features

Low VRAM mode

Low VRAM mode sends Ollama’s low_vram flag with every request, asking the daemon for smaller batches and a leaner KV cache. It trades throughput for the ability to run on hardware that would otherwise OOM.

Where to set it

There are two places this lives:

• Per chat - the Low VRAM toggle in the Performance section of the parameters panel, in both the Simple and Advanced views. Affects just that conversation. Private Chat’s panel has the same toggle.
• Globally - Settings → Features → Low VRAM mode, also offered in setup’s Pick your defaults step, where it starts off. Forces low_vram: true on every Ollama request, Private Chat and model preloading included.

Which one wins

The global setting overrides the per-chat toggle. When the global pin is on, the per-chat toggle shows as on and disabled, with a pointer back to Settings → Features so you understand why it cannot be flipped from the panel.

Turn the global setting off and you regain per-chat control - the pin is never written into a chat’s own parameters, so each chat goes back to whatever it had.

Ollama only

The flag is meaningful only for the local Ollama backend. Loach does not send it to OpenAI-compatible endpoints - hosted services have no concept of your VRAM budget - and in a chat on one, the toggle is disabled with “Ignored by OpenAI providers.”

Still running out of memory

Low VRAM mode is one lever among several. A smaller Context Length, fewer GPU Layers (Advanced view) or a smaller quantization usually save more - see Performance & VRAM.