AI Daily Brief — 12 April 2026

Sunday’s open-source headline: MiniMax drops M2.7, an open-weight self-evolving agent model trained with a self-play loop where the model iteratively improves its own responses. Meanwhile, Ollama ships v0.20.6 to fix crashes seen with Gemma 4 and GLM-5.1.

Top stories

• MiniMax M2.7. One of the first openly-weighted self-evolving agent models, with a self-play training loop for iterative response improvement. via Fazm
• Ollama v0.20.6. The runtime patches crashes when running Gemma 4 and GLM-5.1; a community patch ships for the Gemma 4 26B MoE quantization that had been producing degraded outputs. via DevFlokers

Quiet corners

The first fully AI-authored paper accepted at a top venue circulates over the weekend — a milestone in research automation.

By the numbers

• Self-evolving open-weight models on the table: MiniMax M2.7 joins the small list

Compiled by AI Feed’s editor from publicly reported announcements on 12 April 2026.