Question asked: which open-weight AI models have a permissive license (MIT or Apache-2.0), more than 30B total parameters, a release date after January 2026, and a reported GPQA score — with all four attributes listed per model?
Covers 11 model variants released 2026-02-16 to 2026-08-19; parameter scale in billions of total parameters, GPQA reported as published (mostly GPQA Diamond where named).
| Model | Parameters | Release | License | Reported GPQA | Sources | Link |
|---|---|---|---|---|---|---|
| Qwen3.5-397B-A17B | 397B total / 17B active | 2026-02-16 | Apache-2.0 | 88.4% (GPQA Diamond) | 1 | ai-tldr.dev |
| GLM-5 | 744B total parameters with 40B active per token | 2026-02-17 | MIT | 86.0% on GPQA Diamond | 1 | smartchunks.com |
| Gemma 4 31B Dense | 31B | 2026-04-02 | Apache 2.0 | 84.3% | 4 | huggingface.co |
| GLM-5.1 | 754B total parameters (40B active per token) | 2026-04-07 | MIT license | 86.2% | 3 | morphllm.com |
| Qwen3.6-35B-A3B | 35 billion total, 3 billion active | 2026-04-16 | Apache 2.0 | 86.0 | 2 | labellerr.com |
| DeepSeek V4-Flash | 284B (13B active) | 2026-04-24 | MIT | 88.1 | 1 | thundercompute.com |
| DeepSeek V4-Pro | 1.6T (49B active) | 2026-04-24 | MIT | 90.1 | 2 | thundercompute.com |
| GLM-5.2 | 744B MoE architecture with 40B active parameters | 2026-06-13 | MIT | 91.2 | 8 | groundy.com |
| Inkling | 975B total, 41B active | 2026-07-15 | Apache 2.0 | 87.2% GPQA Diamond | 4 | runaihome.com |
| Inkling-Small | 276B | 2026-07-30 | Apache 2.0 | 89.5% | 1 | developersdigest.tech |
| Ornith-1.5-397B | ~397B | 2026-08-19 | MIT | 92.8 GPQA Diamond | 1 | llm-releases.com |
Live-web scan: 2,766 unique web results searched in a focused verification pass as of 2026-08-31; 11 qualifying rows after deduplicating Gemma 4 and DeepSeek V4-Pro/Pro Max naming. Qualifying means weights under standard MIT or Apache-2.0, strictly over 30B total parameters, released after 2026-01-31, with a reported GPQA score; custom, community, noncommercial and unspecified licenses excluded. Scores are quoted verbatim and may mix GPQA variants and evaluation settings — no ranking is implied. Entries with corroborating-source count 1 rest on a single located page.