The inference framework behind DeepSeek's production serving, built on vLLM, with FlashMLA, DeepGEMM and DeepEP kernels. Apache-2.0.
Why it matters: A rare look at a frontier lab's serving stack; supports V4.1-Flash on SM100 GPUs and Ascend 950.
Agent skill for deep research on DeepSeek's new Search API, with subagents, fact-checking and an HTML report.
Why it matters: Built the same day the search API appeared; a quick way to try it from Claude Code or Codex.
One 8 KB skill that replaces the Reverse Engineer Anything toolkit with plain CLI tools and no MCP.
Why it matters: Claims 38% faster and 35% cheaper runs, a useful data point in the skills-versus-MCP argument.
Peer-to-peer engine, built on iroh, that lets a team's coding agents talk to each other.
Why it matters: Agent-to-agent coordination without a central server, exposed over MCP.
Version control designed for AI coding agents.
Why it matters: One of several attempts to rethink git around agents that commit far more often than humans.
Agent skills for designing, building and testing iPhone Duo and iOS 27 apps, with a fold linter.
Why it matters: Fresh in the agent-skills topic, and an early way to target Apple's foldable.
A neat counterpoint to today's lead.