Qwen3 Models Make Local LLMs Viable as Free Cloud Tiers Disappear

A developer revisiting local large language model performance six months later finds the landscape dramatically changed, with Alibaba's new Qwen3 model lineup delivering usable speed and accuracy on consumer-grade hardware. The Qwen3.6-27B dense model, requiring 32 GB VRAM, is claimed to match Claude 4.5 Opus in accuracy, while the smaller MoE variant Qwen3.6-35B-A3B offers fast performance for lighter tasks. Meanwhile, the Qwen-Coder-Next-80B model targets coding use cases with accuracy comparable to DeepSeek-V3.2 and Kimi K2.5. On the infrastructure side, llama.cpp has introduced an experimental router mode that handles model loading and unloading natively, reducing the need for third-party tools like llama-swap. The shift comes as most free-tier cloud inference providers have either disappeared or become too rate-limited and limited in capability to be practical.