set OLLAMA_NUM_CTX=32768 friday night because a rag demo wanted “just a bit more room”.
activity monitor jumped to ~22gb compressed memory, spotlight beachballed, then the fans sounded like a hairdryer aimed at a metal desk. ollama quit did nothing — had to kill -9 the runner and reboot.
same model at 8k was fine all week. anyone else pinning ctx before they pin hope, or am i the only one still learning this the hard way?
tried Deepseek Harness Desktop this week because the electron wrapper looked less fiddly than wiring ollama myself. was halfway through a prisma schema rewrite on deepseek coder:6.7b local. somehow t…
tried the electron client for DeepSeek Harness this week instead of bouncing between browser tabs. tool call approvals are the part that actually helped — caught a rm rf on a mis resolved path before…
Friday night prompt was literally "just clean up the webhook handler, don't change behavior." It extracted the Stripe signature check into a helper (fine), then quietly deleted the Redis idempotency…
2 comments
Join the discussion
Log in to comment.
yes. default context windows silently kill demos. i keep sticky note: qwen2.5:14b-q4_K_M → num_ctx 8192 on MBP, 16384 only if nothing else open.
32k on 14b is almost always swap theater. sorry for your spotlight.
Activity Monitor lies until Metal owns half the machine. Same story on my Mac Studio —
ollama stopleft a zombie runner twice last month.14b + tight context beats half-loaded giants. I stopped pretending 32k was free after it ate a whole Saturday of evals.