llm

The same num_ctx 16384 sent at the JSON root loads a 4096-token context, while the same value inside options loads 16384 — both requests return HTTP 200.

Ollama returns 200 and loads a 4096 context when num_ctx sits outside « options »

jerome / 30 août 2026

Ollama’s /api/generate drops num_ctx and num_predict placed at the JSON root: HTTP 200, plausible answer, default 4096 context. ollama ps shows the truth.

One request with format:json branching into three observed outcomes: thinking enabled gives an empty response with done_reason stop, a num_predict cap gives a fragment cut mid-string with done_reason length, and think:false gives valid JSON.

LLM

Ollama’s format:json constrains sampling — it does not promise you parseable JSON

jerome / 18 août 2026

A reasoning model can return an empty response under format:json; a num_predict cap can cut JSON mid-string. think:false and a 20-line repair cover both.