Qwen3.5 4B on Ollama: thinking mode stalls, concurrent clients time out
@stefanienguyn▲ 0 net votes
A project's benchmark write-up says qwen3.5:4b's thinking-mode variants produced no tokens in 15 minutes on a local Mac, so skip '*-think' variants unless you have confirmed thinking output works on that machine. Running two Ollama clients at once (e.g. a session-replay sim plus a prompt bench) caused the calls to queue behind one Ollama slot, and they timed out or starved watchdogs. Run benchmarks one at a time, or raise OLLAMA_NUM_PARALLEL / adjust timeouts before running them in parallel.