Edit Issue - .issue.db

Title *

Description

Benchmarked Ollama Cloud models for the [llm.routes] agent route (controlled: identical prompt, num_predict=256, via ollama.com). glm-5:cloud: median 2.5s, ~104 tok/s, consistent. qwen3.5:397b-cloud (prev default): median 5.0s, ~51 tok/s. nemotron-3-ultra:cloud: median 13.5s, ~19 tok/s + highly variable (NOT a win, despite the claim). glm-4.6/glm-4.5:cloud unavailable (410/404). Changed docker/config.ini [llm.routes] agent = ollama/glm-5:cloud, ollama/qwen3.5:397b-cloud, ollama/minimax-m3:cloud (glm-5 primary; qwen+minimax kept as ordered fallbacks). config.ini is mounted -> restarted worker/activity-worker/internal-worker/api (no rebuild). Verified live: worker logs show 'provider=ollama, model=glm-5:cloud'; run_goal '2+2' completed cleanly. LESSON: the apparent speed of a model is dominated by VERBOSITY, not raw tok/s -- qwen3.5:397b over-generated ~8000 tokens on a ~180-word ask (100-200s/call uncapped), which made nemotron *look* faster when it was just less rambly. Always benchmark with num_predict capped for a fair per-token comparison, and consider capping the agent's output tokens to speed up turns regardless of model.

Due Date (YYYY-MM-DD)

Tags (comma separated)

Related Issues (IDs)

Enter IDs of issues related to this one. They will be linked as 'related'.

Edit Issue #786