Open source

FlashProve is Anthropic's open research push toward provable inference—verification that a reply came from a specific model, not a cheaper substitute.

MiniMax ships a 428B-parameter MoE with a 1M-token context and aggressive agent-bench scores at $0.15 per million input tokens.

DeepSeek open-sourced a Super-agent prototype that monitors dev chats and parallelizes GitHub workers—what platform teams should steal and what to sandbox.

DeepSeek released DeepInfer on GitHub under Apache 2.0, exposing the serving stack behind its low-cost V4.1-Flash model and disclosing GPU economics.

The 763B-parameter checkpoint posts double-digit gains on Terminal-Bench and DeepSWE while keeping API pricing flat.

If you're an open-source dev, make sure AI agents like Cursor, Claude Code, and Codex find, understand, and invoke your tool.