Six post-January 2026 open-weight models meet the permissive-license, size and GPQA screen

A live-web scan found six reported matches with more than 30B total parameters, a release after January 2026, a permissive weights license and a published GPQA result. GLM-5.2 leads the reported scores at 91.2 on GPQA-Diamond.

6
models passing all four filters
91.2
highest reported GPQA score
1.6T
largest total parameter count
3
permissive license families

Reported GPQA performance

GPQA score · higher is better
GLM-5.2
91.2
DeepSeek V4 Pro
~90
Qwen 3.5
88.4
Inkling
87.2
GLM-5
~78
Nemotron 3 Ultra
71.9
0255075100

What stands out

01
GLM-5.2 has the top reported result. Its 91.2 GPQA-Diamond score edges the roughly 90 reported for DeepSeek V4 Pro. groundy.com
02
MIT covers half the shortlist. GLM-5, DeepSeek V4 Pro and GLM-5.2 are reported under MIT terms; Qwen 3.5 and Inkling use Apache 2.0.
03
Total size and active size differ sharply. The MoE models activate only 17B–55B parameters per token despite total sizes from 397B to 1.6T.
04
Benchmark comparability is imperfect. Five sources identify GPQA-Diamond; the Qwen 3.5 source reports 88.4 without naming the GPQA variant.

Three leaders

best reported GPQA
GLM-5.2
91.2 GPQA-Diamond. A 753B-total MoE with about 40B active parameters and an MIT license.
largest parameter count
DeepSeek V4 Pro
1.6T total and 49B active parameters. Reported around 90 on GPQA-Diamond under MIT.
strongest apache 2.0 result
Qwen 3.5
397B total and 17B active. Its source reports 88.4 on GPQA but does not identify the variant.

All qualifying models

ModelReleaseParametersLicenseReported GPQASource
GLM-52026-02-13744B total; 40B activeMIT~78% · Diamondaiproductivity.ai
Qwen 3.52026-03-01397B total; 17B activeApache 2.088.4 · variant unstatedintelligibberish.com
DeepSeek V4 Pro2026-04-241.6T total; 49B activeMIT~90 · Diamondaitechconnect.in
Nemotron 3 Ultra2026-06-04550B total; 55B activeOpenMDW-1.171.9% · Diamondmedium.com
GLM-5.22026-06-13753B total; ~40B activeMIT91.2 · Diamondgroundy.com
Inkling2026-07-15975B totalApache 2.087.2% · Diamondkie.ai

Method: live-web search on 26 Aug 2026; 1,242 unique pages scanned in the broad candidate search, with follow-up checks across additional search results. “More than 30B” means total model parameters, not active parameters per token. Scores reproduce the cited reports and are not normalized across prompting, tools or evaluation settings. “Permissive” includes MIT, Apache 2.0 and the reported OpenMDW-1.1 weights license.

Keenable · made with SELECT* · 3,060 pages in 3m 28s · Open the chatShare: XLinkedInReddit