4 items
I benchmarked Qwen3.6-35B-A3B with CPU offloading against Qwopus-9B on 26 coding tasks. The MoE model has 3 billion active parameters — fewer than the dense 9B. It still got crushed. Here is why.
When Anthropic's Claude Code started consuming tokens 4x faster than expected due to a bug, and peak-hour throttling made paid subscriptions unreliable — I realized it was time to stop depending on a single vendor.
From a broken skill evaluation at ADEO to a clean sandbox environment for AI tools — how a DevContainer became my go-to setup for experimenting with AI agents without breaking anything.
Masterize your privacy by running Moltbot locally