Closed models look highly reproducible, while open weights still lag badly
zakkohane · x · 2026-07-23
The follow-up result is blunt: reproducibility looks solved for the closed models, but not for open weights.
- Closed/commercial models cluster high: GPT-5.2 0.94, Claude Opus 4.8 0.92, Grok 4.5 0.89, Gemini 2.5 Pro 0.85.
- Open-weight models remain much less stable: Kimi K2-Thinking 0.28, Llama 3.3 70B 0.24, GLM-4.7 0.14.
- The chart marks 0.8 as the “highly reproducible” line; the open-weight models sit well below it.
Related event: Clinical Triage Alignment Test: Closed Models Excel, Open-Source Lags(6 posts)→
More from Models
- User says €200 Anthropic plan no longer feels worth it after repeated model downgrades — evilsocket · 2026-07-23
- Poolside ships Laguna S 2.1, a compact open-weight coding model that outperforms bigger rivals — The Decoder · 2026-07-23
- AI reportedly solved more than two open math problems a day over the past month — AymericRoucher · 2026-07-23
- Anthropic user says a €200 plan keeps downgrading to weaker Opus versions — evilsocket · 2026-07-23
- Alibaba announces Qwen 3.8, a 2.4T-parameter open-weights model — testingcatalog · 2026-07-23
- Qwen 3.8 Max preview reportedly finishes a Three.js scene task faster than rivals — testingcatalog · 2026-07-23