Open-weight model scan · 20 August 2026
Nineteen post-January releases clear the size, license and GPQA filters
A broad web scan was narrowed to open-weight models released after 31 January 2026, with more than 30B total parameters, a permissive license, and a reported GPQA result. The highest reported score in the surviving set is 91.2 on GPQA Diamond for GLM 5.2.
19
Qualifying model releases
91.2
Highest reported GPQA
1.6T
Largest total parameter count
5
License labels represented
Reported GPQA score
GPQA Diamond unless noted · percent / points
0255075100
What stands out
1
GLM 5.2 leads at 91.2. Nex-N2-Pro, Tencent Hy3 and DeepSeek V4 Pro also report at least 90 points.
2
MoE scale dominates the shortlist. Many models have hundreds of billions of total parameters but activate only 3B–49B per token.
3
Apache 2.0 and MIT account for most entries. NVIDIA’s permissive Open Model License is also included; modified or use-restricted licenses were excluded.
4
Scores are not perfectly apples-to-apples. Most are GPQA Diamond, but tool use, thinking mode, precision and test-time engines differ where noted.
The qualifying releases
| Model | Parameters | Released | License | Reported GPQA |
|---|
Method: 3,038 unique search results were scanned across broad and official-source searches on 20 August 2026. “Parameters” means total model parameters, not active parameters; the cutoff is strictly greater than 30B and the release cutoff is strictly after 31 January 2026. Dates and benchmark setups follow the cited reports/model cards; duplicate aliases and quantized variants were consolidated.