Open-weight models are no longer far behind the best proprietary models, with an average gap of about four months since January 2026, according to Epoch AI. Ollama and LM Studio let users download and run models for free on their own computers, with LM Studio operating entirely offline and smaller models generally requiring about 8 GB of video or unified memory. This approach reduces subscription costs, keeps conversations private and allows users to customize models under Apache 2.0 or MIT licenses. Hardware nevertheless remains a major constraint: small models run on typical computers, while Alibaba’s Qwen 3.8-Max, with 2.4 trillion parameters, is far beyond consumer hardware. Cloud services still lead on the hardest tests, although quantization, GPUs with 24 GB of video memory and machines with 80 GB of memory make it possible to run larger models locally.
Source: Read the original article


