Tuskira open-sourced an agent gateway (Apache-2.0, v0.4.0 alpha) that centralizes MCP tools, model-provider credentials/routing, per-agent profiles and policy enforcement, audit logs, and token-cost visibility in a self-hosted deployment. It surfaced in this pass as a new governance/infrastructure category: one control plane for both agent-to-tool and agent-to-model traffic; still pre-1.0, with APIs subject to change. Source: https://www.tuskira.ai/blog/open-source-ai-agent-gateway
Ai2 open-sourced AstaBrief-8B, a Qwen3-8B-derived text model specialized for turning research questions plus retrieved literature into cited reports; weights and training data are available under Apache 2.0, and an example local-PDF workflow is included. It powers Asta’s Fast mode, which Ai2 reports averages 51.1 seconds end-to-end versus 178.5 seconds for its Claude-powered Thinking mode. On ScholarQA-CS2, it scores 87.0 versus 86.2 for Asta ScholarQA; caveat: much evaluation was completed in 2025 and Ai2 has not rerun against current frontier models. The useful agentic angle is a small task-specific model handling evidence-grounded report synthesis locally/quickly within a larger research workflow. Claims are author-reported. Sources: https://huggingface.co/blog/allenai/astabrief ; https://huggingface.co/allenai/AstaBrief_8B