Nine post-January 2026 open-weight models clear a strict permissive-license test

The screen required more than 30B total parameters, weights released after 31 January 2026, a reported GPQA result, and a standard MIT or Apache 2.0 license. GLM-5.2 leads this strict-license set at 91.2 on GPQA-Diamond.

qualifying models

91.2

top GPQA-Diamond score

36B

smallest qualifying model

standard licenses represented

Reported GPQA score, points; bars start at 80 to show separation

GPQA / GPQA-Diamond

GLM-5.2

91.2

Hy3

90.4

DeepSeek V4-Pro

90.1

Darwin-36B-Opus

90.0

LongCat-2.0

88.9

DeepSeek V4-Flash

88.1

GLM-5.1

86.2

GLM-5

86.0

Gemma 4 31B

84.3

80848892

What the screen shows

GLM-5.2 has the highest reported score. Its 91.2 GPQA-Diamond result leads this strict MIT/Apache shortlist. labindex.ai

Size does not determine the ranking. Darwin-36B-Opus reports 90.0, edging models hundreds of billions of parameters larger. theagenttimes.com

DeepSeek contributes two distinct checkpoints. V4-Pro reports 90.1 while the smaller V4-Flash reports 88.1; both are MIT-licensed. thundercompute.com

Custom “modified MIT” licenses are outside this strict count. They may permit commercial use, but added model-specific conditions mean they are not treated as standard permissive licenses here.

Three useful reference points

highest GPQA

GLM-5.2

91.2

753B total parameters, MIT. Open weights dated 16 June 2026.

smallest model

Darwin-36B-Opus

36B

Reports 90.0 GPQA-Diamond under Apache 2.0.

largest model

DeepSeek V4-Pro

1.6T

49B active parameters and a reported 90.1 GPQA-Diamond score.

Qualifying models

Model	Parameters	Release	License	GPQA	Source
GLM-5	744B total; 40B active	2026-02-11	MIT	86.0 Diamond	dev.to
Gemma 4 31B	30.7B–32.2B reported	2026-04-02	Apache 2.0	84.3 Diamond	telnyx.com
GLM-5.1	744B total; 40B active	2026-04-07	MIT	86.2	winbuzzer.com
Darwin-36B-Opus	36B	2026-04-08	Apache 2.0	90.0 Diamond	theagenttimes.com
DeepSeek V4-Flash	284B total; 13B active	2026-04-24	MIT	88.1 Diamond	thundercompute.com
DeepSeek V4-Pro	1.6T total; 49B active	2026-04-24	MIT	90.1 Diamond	thundercompute.com
GLM-5.2	753B total; ~40B active	2026-06-16	MIT	91.2 Diamond	labindex.ai
LongCat-2.0	1.6T total; 48B active	2026-06-30	MIT	88.9 Diamond	lumichats.com
Hy3	295B total; 21B active	2026-07-06	Apache 2.0	90.4 Diamond	ai-beat.github.io

Method: live-web research on 25 August 2026; three broad search passes using 24 query formulations, with final evidence drawn from 9 pages across 9 sites. “Parameters” means total model parameters, not active parameters; “permissive” is restricted to unmodified MIT or Apache 2.0 model-weight licenses. Scores are reported figures, not independently rerun evaluations, and GPQA settings may differ by publisher.