Ten permissive open-weight models over 30B now report GPQA, and three cluster above 90

A live-web inventory of 10 qualifying model releases — permissive license, more than 30B total parameters, released after January 2026, with a reported GPQA score — spanning February through July 2026, scanned through 2026-08-27. Scores are shown exactly as reported; evaluation conditions may differ.

Darwin-36B-Opus reaches 90.0% on GPQA Diamond with only 36B parameters — the smallest of the three 90-plus models by more than an order of magnitude — theagenttimes.com

Kimi K2.6 leads the field at 90.5 with 1000B total parameters and only 32B active, just ahead of DeepSeek-V4-Pro's 90.1% at 1.6T total / 49B active — deepinfra.com, awesomeagents.ai

Nemotron 3 Ultra is the low outlier at 71.9% despite 550 billion parameters — 12.4 points under the next-lowest score in the set — medium.com

Apache 2.0 covers six of the ten releases; NVIDIA's NVFP4 V2 quantization of Qwen3.5-397B-A17B holds 87.7 versus 87.2 for FP8 — huggingface.co

All 10 qualifying releases, sorted by release date

Model	Released	Parameters	License	Reported GPQA	Source

Live-web inventory of open-weight AI model releases: permissive license, >30B total parameters (active MoE parameters shown as supplementary), released after January 2026, with a reported GPQA score. 10 rows; scan cutoff 2026-08-27. GPQA scores are transcribed exactly as reported — mostly GPQA Diamond — and normalized to a percentage for the chart; evaluation conditions may differ across sources, and some entries rely on secondary reporting, so this is an inventory, not a definitive registry.