FULL STORY

Deep Dive into Kimi K3 Architecture and Agent Practices

Moonshot AI and Together AI hosted a webinar alongside Hugging Face, deeply analyzing the underlying architecture and agent practices of the newly open-sourced Kimi K3 model.

2026-07-28 ~ 2026-07-30 · 2 episodes · 9 posts

Episode 1 · Together AI and Moonshot to Share Kimi K3 Architecture and Agent Practices (2026-07-28, 6 posts)

Together AI and Moonshot AI have announced a joint technical webinar on the Kimi K3 model, scheduled for July 30 at 9:00 AM PDT. The session will be led by Moonshot engineer Feihu Tang, focusing on the model's architectural design and its practical application in production-grade agents.

Confirmed

  • Event details: Scheduled for July 30 at 9:00 AM PDT, featuring a presentation by Moonshot engineer Feihu Tang.
  • Model specs: According to the announcement, Kimi K3 is the first open 3T-c parameter model (some promotions note 2.8T) supporting a context length of up to 1 million tokens.
  • Architecture insights: The event will provide a deep dive into the underlying architecture and design trade-offs of Kimi K3.
  • Agent workflow: The core focus will be on the challenges of long-running agent workflows, specifically addressing context persistence across code, tools, vision, and research, while sharing best practices for deploying Kimi K3 in production environments.

Why it matters

As large language models increasingly pivot towards agentic workflows, maintaining context and ensuring tool stability during long-running tasks have become major pain points. This webinar offers both a technical look under the hood of Kimi K3 and practical guidance for developers building robust, production-grade agents.

Episode 2 · Tech Communities Deep Dive into Kimi K3 Model (2026-07-30, 3 posts)

Hugging Face and Together AI hosted in-depth discussions with Moonshot AI, analyzing the architecture, technical details, and applications of the newly open-sourced Kimi K3 model based on its technical report.