Local models run out of room faster than you expect, and that is the most common cause of a bad answer. Giving them more room matters more than anything else here.
The Ollama catch. It reports the architecture's maximum context, not the window it is actually serving. qwen3.5 reports 262144 and is served 4096.
Because of that, only an explicit num_ctx is trusted — a model without one is reported as unknown rather than guessed at. Otherwise the ladder would happily "upgrade" from an 8192 model to one it believed had 262144 and that actually had 4096.
Create a variant with the window you want, and it appears on the ladder with a real limit:
$ printf 'FROM qwen3.5:latest\nPARAMETER num_ctx 32768\n' \ > Modelfile $ ollama create qwen3.5-32k -f Modelfile
Then point default at it, or declare the window in the config if you set it server-wide instead.
The status row carries a context bar. When it turns amber, /clear starts fresh — and /status says whether any rung on the ladder is actually roomier than what you are on.
auto · ⎇ main · qwen3.5-8k:latest · ██░░░░░░░░ 22% · 1.8kAnything served from your own machine joins the escalation ladder without needing pricing to say so. That is what makes the local-first arrangement work: most turns cost nothing, and a hosted rung is there only for the ones that genuinely outgrow the window. See what it costs.