Local LLM choices shift from benchmarks to memory limits

Open-weight LLM selection in September 2026 is increasingly shaped by what hardware can actually run. Chinese models now dominate key usage signals: by May 2026, they accounted for roughly 61% of tokens routed through OpenRouter, while Qwen crossed one billion cumulative downloads on Hugging Face and passed Llama. More than 200,000 Qwen-tagged models are on the hub, and roughly 40% of new LLM derivatives there are built on a Qwen base.

The most practical local tier is the workstation class. Qwen3.8-27B in Q4_K_M fits inside an RTX 3090 or 4090, supports 262K context, carries an Apache 2.0 license and ranks close to much larger systems. Laptops can handle smaller models such as Qwen3 8B or Gemma 4 E4B, but context length and KV cache growth can make nominally fitting models fail mid-task. Larger sparse MoE systems such as Kimi K3 and Qwen3.8-Max remain rack-scale deployments despite open weights.

Cloud pricing complicates the local hardware case. A $20,000 Mac Studio would need more than fourteen years of continuous generation at 10 tok/s to match the output tokens that the same money buys in GLM-5.3 cloud usage. Local machines make more sense for privacy, offline work and control over model behavior than for simple cost savings.