ran qwen3:14b through Ollama on the beat-up M2 last night. at num_ctx 8192 it edits a small repo and calls tools normally; at 16k it starts repeating the tool schema, then spends 40 seconds apologising to itself.
I switched back to 8k and added a tiny summary after each file. less clever, but it actually finished the task. wondering if anyone has a local setup that stays sane past 12k without turning the Mac into a space heater.
Asked Claude to speed up local LLM evals on my M2. It rewrote conftest.py , marked 14 flaky tests with @pytest.mark.skip(reason="ollama timeout") , and CI finished in 41 seconds. The skip reason is a…
MacBook Air M2, 8GB. started a long refactor prompt in the desktop client around 2pm at a cafe. about 12 minutes in the app just quit and relaunched into an update screen. draft gone. Progress bar sa…
was mid refactor in Cursor last night with qwen2.5 14b via ollama. replies felt sharp for 25 minutes, then answers got weirdly polite and slow. checked the bar: still said qwen. ran ollama ps — it ha…
4 comments
Join the discussion
Log in to comment.
i hit the same wall with a LangGraph loop, but my culprit was the tool descriptions, not qwen. the summary node fixed it until it started eating 600 tokens every pass. check the prompt after 3-4 steps, not just the final length — Retry-After for context, basically.
repro question: are you injecting the MCP schemas on every turn? i saw qwen3:14b behave fine at 12k until one server exposed 31 tools; the model was okay, the prompt was just mostly JSON wearing a trench coat.
yeah the tool dump is the silent killer. i strip MCP schemas to name + one-line desc before every turn; full JSON only when the model actually picks a tool. qwen3:14b stopped echoing schemas after that. still heats the laptop though.
same on a 16gb m2. once i bumped num_ctx past 12k ollama's rss went from ~9gb to thrashing and the fan sounded like a leaf blower. capped ctx at 10k and stuffed a rolling 800-token summary into the system prompt. ugly but it stopped the apology loop.
are you on metal or cpu?