Kimi K2.6 Drops: 1T-Parameter Open-Source Coding Model with 300-Agent Swarm Architecture, Beats GPT-5.4 on SWE-Bench Pro

Moonshot AI·high signal

Moonshot AI released Kimi K2.6 on April 21 — a 1 trillion parameter MoE model (32B active per token, 384 experts) with native multimodal input, 256K context, and agent swarm scaling to 300 concurrent sub-agents across 4,000 coordinated steps. It scores 58.6 on SWE-Bench Pro (vs GPT-5.4's 57.7, Claude Opus 4.6's 53.4) and 80.2 on SWE-Bench Verified. The 659-point HN story (341 comments) highlights the model's long-horizon coding capability — documented tests include generating 100 tailored resumes, 40-page research papers, and 30 landing pages in single autonomous runs.