The state of open source AI
Points and comments are a snapshot, not live.
Open-source AI models have reached parity with closed models on most tasks, but deployment remains harder.
Open-weight models now handle a majority of production tokens on OpenRouter, with the five highest-volume models all open. The capability gap on Chatbot Arena dropped from 8% to 0.5% by August 2024 before reopening to 3.3% as closed reasoning models advanced. Inference costs for GPT-4-class models fell 50x in 36 months ($20 to $0.40 per 1M tokens). However, only 51% of open-model teams reach production versus 63% for closed, held back by operational tooling and trust gaps rather than model capability. Chinese open-weight models dominate downloads, with Alibaba's Qwen at 942M Hugging Face downloads versus Meta's Llama at 476M.
What commenters are saying
Many commenters criticized the article's presentation and content. Several found the scrolling reveal animations inaccessible and the custom fonts hard to read. One commenter called the analysis "slop" and questioned the data visualization's accuracy, noting bars that didn't align with labels on mobile. A more substantive debate emerged: some argued open models will kill Anthropic and OpenAI, as hyperscalers can run them without licensing fees and the frontier advantage shrinks. Others countered that model quality still varies significantly and that training open models requires expensive compute from organizations willing to absorb zero return, which may not persist.