Twelve post-January releases clear 30B parameters, permissive licensing and a reported GPQA score
A live-web scan found 12 distinct model releases matching the stated filters when Modified MIT licenses that retain commercial use are counted as permissive. Hy3 leads the reported GPQA results at 90.4; evaluation settings differ, so the scores are reported rather than treated as a controlled leaderboard.
12
qualifying models
90.4
highest reported GPQA
31B
smallest total size
3
license families
Reported GPQA score
GPQA Diamond, %
Hy3
90.4
GLM-5.2
89.39
DeepSeek-V4-Flash
89.1
LongCat-2.0
88.9
Qwen3.5-397B-A17B
88.4
Kimi K2.5
87.6
Inkling
87.2
Qwen3.5-122B-A10B
86.6
Gemma 4 31B
84.3
Qwen3.5-35B-A3B
84.2
K-EXAONE 2.0
82.2
Mistral Medium 3.5
76.88
0255075100%
What stands out
1
Hy3 reports the top score. Tencent’s 295B-total model reports 90.4 on GPQA Diamond under Apache 2.0. huggingface.co
2
Mixture-of-experts models dominate. Nine of the twelve report a smaller activated-parameter count than their total size, including Kimi K2.5 at 1T total and 32B active. huggingface.co
3
Apache 2.0 is the most common license. Six releases use Apache 2.0, four use MIT, and two use a Modified MIT license. arxiv.org
Qualifying releases
| Model | Released | Parameters | License | Reported GPQA | Source |
|---|---|---|---|---|---|
| Kimi K2.5 | 2026-02-02 | 1T total; 32B active | Modified MIT | 87.6, Diamond, thinking avg@8 | HF |
| Qwen3.5-397B-A17B | 2026-02-16 | 397B total; 17B active | Apache 2.0 | 88.4, Diamond | HF |
| Qwen3.5-122B-A10B | 2026-02-24 | 122B total; 10B active | Apache 2.0 | 86.6, Diamond | HF |
| Qwen3.5-35B-A3B | 2026-02-24 | 35B total; 3B active | Apache 2.0 | 84.2 | HF |
| Gemma 4 31B | 2026-04-02 | 31B dense | Apache 2.0 | 84.3, Diamond | HF |
| DeepSeek-V4-Flash | 2026-04-24 | 284B total; 13B active | MIT | 89.1, Diamond | HF |
| GLM-5.2 | 2026-06-16 | 753B total; 40B active | MIT | 89.39, Diamond | HF |
| Mistral Medium 3.5 128B | 2026-07-01 | 128B | Modified MIT | 76.88 FP8 / 76.80 NVFP4, Diamond | HF |
| LongCat-2.0 | 2026-07-02 | 1.6T total; 48B active | MIT | 88.9, Diamond | HF |
| Hy3 | 2026-07-06 | 295B total; 21B active | Apache 2.0 | 90.4, Diamond | HF |
| Inkling | 2026-07-15 | 975B total; 41B active | Apache 2.0 | 87.2, Diamond | HF |
| K-EXAONE 2.0 | 2026-08-05 | 750B total; 37B active | Apache 2.0 | 82.2, Diamond | arXiv |
Method: web search conducted 2026-08-24 across 2,190 unique result pages; 12 canonical model releases survived checks for release after 2026-01-31, strictly more than 30B total parameters, commercial-use permissive licensing, and a reported GPQA score. Scores are publisher-reported and may use different inference settings. Modified MIT is included where commercial use and modification remain permitted.