7 Open-Weight AI Models Released After Gemma 4 That Self-Hosters Should Know in 2026

Following Google's Gemma 4 release in late March 2026, at least seven new open-weight models from major AI labs appeared within six months, many surpassing Gemma 4 on key benchmarks. Meta's Muse Glimmer 30B, released on August 10 under a full Apache 2.0 license, topped agent-task benchmarks including MCP Atlas at 75.5 versus Gemma 4's 54.2, marking Meta's return to fully open model weights. Alibaba continued its steady release cadence with Qwen3.6-27B in April and Qwen3.8-27B in August, both reportedly outperforming Gemma 4 31B across general knowledge, math, and coding tasks. Most of the seven models fit within a standard 24 GB GPU, making them accessible to individual self-hosters, while larger models like Kimi K3 and GLM-5.3-Flash require server-grade hardware. The rapid succession of releases highlights a trend where open-weight models are now being refreshed every three to four months, quickly dating even recently strong baselines like Gemma 4.