Eleven permissive open-weight models over 30B report GPQA scores, all released February–August 2026

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).

The highest reported score, 92.8 on GPQA Diamond, belongs to Ornith-1.5-397B (MIT, ~397B, released 2026-08-19) — but rests on a single located source. llm-releases.com
GLM-5.2 pairs a 91.2 score with the strongest corroboration in the set: 8 independent pages report it. groundy.com
Parameter scale spans 52× — from Gemma 4 31B Dense, the only model near the 30B cutoff, to DeepSeek V4-Pro at 1.6T total (49B active). thundercompute.com
Licenses split 6 MIT to 5 Apache-2.0; 4 of the 11 variants have only one located reporting source, so their figures are unconfirmed. huggingface.co

All 11 qualifying models

ModelParametersReleaseLicenseReported GPQASourcesLink
Qwen3.5-397B-A17B397B total / 17B active2026-02-16Apache-2.088.4% (GPQA Diamond)1ai-tldr.dev
GLM-5744B total parameters with 40B active per token2026-02-17MIT86.0% on GPQA Diamond1smartchunks.com
Gemma 4 31B Dense31B2026-04-02Apache 2.084.3%4huggingface.co
GLM-5.1754B total parameters (40B active per token)2026-04-07MIT license86.2%3morphllm.com
Qwen3.6-35B-A3B35 billion total, 3 billion active2026-04-16Apache 2.086.02labellerr.com
DeepSeek V4-Flash284B (13B active)2026-04-24MIT88.11thundercompute.com
DeepSeek V4-Pro1.6T (49B active)2026-04-24MIT90.12thundercompute.com
GLM-5.2744B MoE architecture with 40B active parameters2026-06-13MIT91.28groundy.com
Inkling975B total, 41B active2026-07-15Apache 2.087.2% GPQA Diamond4runaihome.com
Inkling-Small276B2026-07-30Apache 2.089.5%1developersdigest.tech
Ornith-1.5-397B~397B2026-08-19MIT92.8 GPQA Diamond1llm-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.

Keenable · made with SELECT · 4,429 pages in 4m 08s · Ask your own questionShare: XLinkedInReddit