Twenty-one post-January releases pair permissive weights with a reported GPQA score

A broad web scan found 21 distinct models meeting the strict filters after excluding Kimi K2.5, whose primary model card dates its release to January 29. Reported GPQA results range from 67.9 to 91.2, but differing variants and evaluation setups mean the scores are not a controlled leaderboard.

qualifying models

91.2

highest reported GPQA

Apache 2.0 releases

MIT releases

Reported GPQA score

Points or percent; GPQA Diamond unless source labels only GPQA

blue = vendor or cited reported score

GLM-5.2

91.2

Nex-N2-Pro

90.7

Tencent Hy3

90.4

MiMo V2.5-Pro

90.2

DeepSeek V4 Pro

90.1

Darwin-36B-Opus

90.0

Inkling Small

89.5

Qwen 3.5 397B

88.4

DeepSeek V4 Flash

88.1

Inkling

87.2

Qwen 3.5 122B

86.6

GLM-5.1

86.2

Qwen 3.6 35B

86.0

Gemma 4 31B

84.3

Qwen 3.5 35B

84.2

Motif 3

83.4

K-EXAONE 2.0

82.2

GLM-5

~78

Command A+

~76

Mistral Small 4

70.0

QwQ-32B-v2

67.9

0255075100

What stands out

GLM-5.2 has the highest reported result. Its MIT-licensed 753B MoE reports 91.2 on GPQA-Diamond. huggingface.co

Apache 2.0 is the dominant permissive license. Fourteen of the 21 qualifying models use Apache 2.0; seven use MIT.

Most are sparse MoEs. Many have hundreds of billions of total parameters but activate only 3B–49B per token, so total size is not inference cost.

The cutoff matters. Kimi K2.5 was excluded because its own model-card changelog gives January 29, 2026, despite later pages assigning it February or April dates. huggingface.co

Qualifying models

Model	Release	Parameters	License	Reported GPQA
Qwen 3.5 122B-A10B	2026-02-01	122B / 10B active	Apache 2.0	86.6 Diamond · source
Qwen 3.5 35B-A3B	2026-02-01	35B / 3B active	Apache 2.0	84.2 Diamond · source
Qwen 3.5 397B-A17B	2026-02-01	397B / 17B active	Apache 2.0	88.4 · source
GLM-5	2026-02-11	744B / 40B active	MIT	~78 Diamond · source
Mistral Small 4	2026-03-16	119B / 6.5B active	Apache 2.0	0.70 reported (70%) · source
QwQ-32B-v2	2026-03-16	32B	Apache 2.0	67.9% · source
Gemma 4 31B	2026-04-02	31B	Apache 2.0	84.3 Diamond · source
GLM-5.1	2026-04-07	754B / 40B active	MIT	86.2 Diamond · source
Darwin-36B-Opus	2026-04-08	36B	Apache 2.0	90.0 Diamond · source
Qwen 3.6 35B-A3B	2026-04-16	35B / 3B active	Apache 2.0	86.0 Diamond · source
DeepSeek V4 Flash	2026-04-24	284B / 13B active	MIT	88.1 Diamond · source
DeepSeek V4 Pro	2026-04-24	1.6T / 49B active	MIT	90.1 Diamond · source
Command A+	2026-05-20	218B / 25B active	Apache 2.0	~76 Diamond · source
Nex-N2-Pro	2026-06-02	397B / 17B active	Apache 2.0	90.7 Diamond · source
Xiaomi MiMo V2.5-Pro 1T-A42B	2026-06-03	1T / 42B active	MIT	90.2 Diamond · source
GLM-5.2	2026-06-13	753B / 40B active	MIT	91.2 Diamond · source
Tencent Hy3 295B-A21B	2026-07-06	295B / 21B active	Apache 2.0	90.4 Diamond · source
Inkling	2026-07-15	975B / 41B active	Apache 2.0	87.2 Diamond · source
Inkling Small	2026-07-30	276B / 12B active	Apache 2.0	89.5% · source
K-EXAONE 2.0	2026-08-05	750B / 37B active	Apache 2.0	82.2 Diamond · source
Motif 3	2026-08-18	314B / 13.2B active	MIT	83.4 Diamond · source

Method: broad web search across 2,472 deduplicated pages, checked 2026-08-23. “Parameters” means total model parameters; active parameters are shown for MoE models. Included only dates later than 2026-01-31 and standard permissive Apache 2.0 or MIT licenses. Scores are reported figures, not normalized independent retests.