Open-Weight AI Models Trail Closed Versions by Four Months

Epoch AI's May 29 analysis indicates that open-weight models have trailed the best closed models by an average of four months since January 2026. In Alibaba's internal comparison, closed models won 24 of 31 tests against open models. Despite this performance gap, users can run open-weight models locally using free programs like Ollama and LM Studio. This approach avoids monthly subscriptions, which typically cost about $20, and prevents conversations from being sent to external servers. LM Studio can operate entirely offline after downloading model files.

Running these models locally requires specific hardware. Users generally need at least 8GB of VRAM or Mac unified memory for smaller models. Serious users should purchase a gaming GPU with 24GB of VRAM or more. Quantization helps by shrinking models through fewer bits, such as using 8-bit integers instead of 32-bit floating points. LM Studio recommends 16GB or more of RAM on Apple Silicon Macs running macOS 14 or newer, while Windows users need at least 16GB of RAM and 4GB of dedicated VRAM. Intel-based Macs are not supported by LM Studio.

Several models are available under permissive licenses. DeepSeek's R1 series is under the MIT License, supporting commercial use and modifications. Google's Gemma 4 and OpenAI's gpt-oss ship under the Apache 2.0 license. Ollama's library lists Gemma 4 26B at 16 to 19GB and the 31B version at 19 to 20GB. OpenAI's gpt-oss-20b needs only 16GB of memory, while gpt-oss-120b runs within 80GB. Separately, a court ordered OpenAI to preserve ChatGPT logs, including deleted ones.