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

2026-07-18 · Source: AI Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Moonshot AI has open-sourced Kimi K3, introducing it as the first open 3-trillion-class model, with weights scheduled for release on July 27. This 2.8-trillion-parameter Stable LatentMoE model activates 16 of 896 experts, features a 1M-token context window, and includes native vision capabilities. Kimi Delta Attention enhances decoding speed by up to 6.3x in million-token contexts, while Attention Residuals boost training efficiency by approximately 25% with minimal extra compute. Demonstrating strong agentic performance, Kimi K3 leads Terminal-Bench 2.1 and scores 67.3 on DeepSWE, showcasing autonomous kernel optimization and vision-in-the-loop coding. It is currently available via Kimi.com, Kimi Code, and API, priced from 0.30 dollars to 15 dollars per million tokens.

Key takeaway

For AI Engineers developing agentic systems, Moonshot AI's Kimi K3 offers a powerful new open-source option. Its 2.8T parameters, 1M-token context, and native vision, combined with faster decoding via Kimi Delta Attention, make it suitable for complex tasks like autonomous kernel optimization. Consider integrating Kimi K3 via API now, or plan for its open weights release on July 27 to explore its advanced agentic capabilities for your projects.

Key insights

Kimi K3 introduces a 2.8T-parameter open model with novel attention mechanisms for enhanced efficiency and agentic capabilities.

Principles

• Stable LatentMoE enables massive scale with efficient activation.
• Novel attention designs can significantly boost decoding speed.
• Integrated vision enhances agentic coding tasks.

In practice

• Utilize Kimi K3 for large-scale agentic development.
• Apply Kimi Delta Attention for faster decoding.
• Employ native vision for complex coding tasks.

Topics

• Kimi K3
• Large Language Models
• MoE Architecture
• AI Agents
• Attention Mechanisms
• Open-Source AI
• Vision-Language Models

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Moonshot AI launches Kimi K3 — Kimi K3's release intensifies open model competition, potentially narrowing the performance gap with proprietary LLMs.
• Last Week in AI #334 - Kimi K2.5 & Code, Genie 3, OpenClaw & Moltbook — Multimodal AI models and coding agents are rapidly advancing, alongside new interactive world-building and always-on AI assistants.
• [AINews] Moonshot Kimi K2.6: the world's leading Open Model refreshes to catch up to Opus 4.6 (ahead of DeepSeek v4?) — Chinese labs are rapidly advancing open-weight agentic coding models with strong performance and ecosystem integration.
• 5 agentic micro-products you can ship this weekend with Kimi K3 — Kimi K3 offers agentic capabilities at a cost enabling solo developers to build complex micro-products.
• Kimi K3: Can Moonshot’s New Model Be Actually Used For Free? — Kimi K3 is a 2.8T parameter model with a 1M token context window, specialized for coding and autonomous agents.
• Building and Training a Kimi-K2 Model Using DeepSeek-V3 Components — Kimi-K2 optimizes agentic LLM performance through architectural sparsity, a novel optimizer, and enhanced training data utility.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Newsletter.