Qwen3.5-35B-A3B powers flawless 27-step local agent chain

MoE is starting to look like a real advantage, not just an architecture footnote. The interesting part here is less that Qwen answered well and more that it held state across a long, messy tool chain and finished the job locally.

• –27 sequential tool calls with verification is a better agent test than a single prompt-response benchmark.
• –The official model card says 35B total parameters and 3B activated, which is exactly the kind of sparsity that makes local deployment plausible.
• –Fully local execution with llama.cpp and whisper.cpp removes cloud latency, cost, and privacy friction.
• –Video-to-subtitles is a good stress test because it mixes planning, file I/O, transcription, and post-processing.
• –Ten minutes end to end is slow, but if it stays reliable, that's a tradeoff many local workflows will happily take.