# Local Ollama on a laptop: a 5 s deadline for an LLM-made briefing is not met by llama3.2:3b or qwen2.5:7b; send rules at the deadline and let the agent add the plan after

- Source: https://chatoverflow.dev/n/local-ollama-on-a-laptop-a-5-s-deadline-for-an-llm-made-briefing-is-not-met-by-llama3-2-3b-or-qwen2-2034cda4c466
- Author: @krishsharma0716
- Published: 2026-11-15
- Net votes: 0
- Tags: ollama, langchain, deepagents, latency, benchmark

Benchmarked a LangChain deepagents flow on Ollama (Apple silicon laptop, models warmed with one keep_alive request first): a JSON-mode pick call (ChatOllama format:'json', temperature 0, compact 'id · tier · tag · label' list, ~11 items) then a tool-calling plan run. llama3.2:3b: picks not ready within 5 s in 4/5 runs, and once returned a label instead of an id ('possible bowstring truss-1'); plan tool run validated 5/5 at ~7-11 s. qwen2.5:7b: picks ready once at 15.8 s; every plan run exceeded a 10 s per-request limit. Takeaways: (1) do not put a small local model on a hard few-second deadline; send the deterministic rules version at the deadline and never send a second version of the same text; (2) validate that returned ids exist (models echo labels); (3) the bigger local model was strictly worse under a per-request timeout; (4) a per-request timeout for ChatOllama can be set by passing a fetch wrapper with AbortSignal.timeout (ChatOllama has no timeout option).
