🤖 AI Agents Weekly: Kimi K3, Thinking Machines Inkling, Agentic Misalignment, Perplexity SPACE, Sunday ACT-2, GPT-Red, and More

Moonshot AI open-sourced Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window and native vision, achieving state-of-the-art results on agentic benchmarks like Terminal-Bench 2.1 and DeepSWE. The model introduces Kimi Delta Attention for faster decoding and Attention Residuals for improved training efficiency, with weights to be released on July 27.

🤖 AI Agents Weekly: Kimi K3, Thinking Machines Inkling, Agentic Misalignment, Perplexity SPACE, Sunday ACT-2, GPT-Red, and More Kimi K3, Thinking Machines Inkling, Agentic Misalignment, Perplexity SPACE, Sunday ACT-2, GPT-Red, and More In today's issue: Moonshot open-sources Kimi K3 Thinking Machines ships Inkling weights Anthropic finds new agent misalignment Perplexity opens SPACE agent runtime Sunday Robotics unveils ACT-2 OpenAI launches GPT-Red red teamer Meituan open-sources LongCat-2.0 schema harness saturates ARC-AGI-3 And all the top AI dev news, papers, and tools. Top Stories Moonshot Open-Sources Kimi K3 Moonshot AI introduced Kimi K3, billed as the first open 3-trillion-class model, with weights due July 27. Scale and architecture: A 2.8T-parameter Stable LatentMoE activating 16 of 896 experts, with a 1M-token context window and native vision in a single model. New attention design: Kimi Delta Attention delivers up to 6.3x faster decoding in million-token contexts, and Attention Residuals add roughly 25% higher training efficiency at under 2% extra compute. Agentic strength: Leads Terminal-Bench 2.1 and scores 67.3 on DeepSWE, with demos of autonomous kernel optimization, compiler development, and vision-in-the-loop coding. Availability: Live via Kimi.com, Kimi Code, and API from about 0.30 dollars to 15 dollars per million tokens, with open weights following on July 27.