The gap between open weights LLMs and closed source LLMs (analysis)

What happened

A Doubleword analysis trended on Hacker News on June 26 quantifying how quickly open-weight models are closing the gap with closed-source frontier. The piece argues that the gap on standard benchmarks is now <5%, and in some domains (code, MMLU) open weights actually lead.

Context and impact

It dovetails directly with a same-day CNBC piece that companies are turning to cheaper open-source alternatives. For OpenAI and Anthropic — especially given new federal restrictions — it is a warning sign that enterprise customers may migrate to self-hosted Llama/Qwen/GLM.

Details

• DeepSeek, Qwen 3, GLM 5.2, and Llama 4 cited as competitive
• Inference cost on open weights often 5–10× cheaper
• Enterprise advantages: data privacy, on-prem, fine-tuning control
• Counter-argument: closed source still leads on multi-step agentic tasks
• Reads as a complement to the 'tokenmaxxing → efficiency' shift narrative