🤖 AI Agents Weekly: MiniMax-M2, Cursor 2.0, SWE-1.5, Agent Data Protocol, Kimi CLI

MiniMax-M2, Cursor 2.0, SWE-1.5, Agent Data Protocol, Kimi CLI

In today’s issue:

• MiniMax-M2 releases as a top-tier open-source model

• Cursor introduced Cursor 2.0

• Kimi launches Kimi CLI

• Research introduces proactive context management

• Cognition launches SWE-1.5, a fast frontier-scale coding agent

• Microsoft releases a new RL training framework for AI agents

• Research on how AI Agents work differently from humans

• Agent Data Protocol

• Top AI dev news, tool updates, and more.

---

---

Top Stories

MiniMax-M2

MiniMax-M2 is an open-source mixture-of-experts (MoE) language model optimized for coding and agentic workflows. The model features 230 billion total parameters with only 10 billion active parameters, striking a balance between capability and computational efficiency.

Key highlights:

• Leading performance – MiniMax-M2 ranks #1 among open-source models and #5 overall on Artificial Analysis, excelling at multi-file code edits, test-validated repairs, and compile-run-fix loops. Also achieves #1 token usage among open-source models on OpenRouter.

• Strong agent capabilities – Demonstrates robust browser automation and information retrieval in agent scenarios, with an interleaved thinking model that maintains reasoning in <think>...</think> tags and tool-calling capabilities for complex agent workflows.

• Benchmark results – Achieves 69.4% on SWE-bench Verified, 46.3% on Terminal-Bench (outperforming Claude Sonnet 4’s 36.4%), 44% on BrowseComp (substantially above Claude Sonnet 4.5’s 19.6%), 36.2% on Multi-SWE-Bench, and 75.7% on GAIA (text-only).

• Efficient architecture – The 10B activation footprint (from 230B total parameters) enables faster feedback cycles in agentic loops, higher throughput for concurrent runs, and simpler deployment infrastructure, addressing practical pain points in production agent systems.

• Availability – Open-source weights available via HuggingFace, free API access (limited time) on the MiniMax platform, MiniMax Agent product at agent.minimax.io, and supports SGLang and vLLM for local deployment.

GitHub