Cover Image for How advanced tool calling transforms agentic use cases: A conversation with Moonshot AI
Cover Image for How advanced tool calling transforms agentic use cases: A conversation with Moonshot AI
Avatar for Together AI Calendar
Hosted By

How advanced tool calling transforms agentic use cases: A conversation with Moonshot AI

YouTube
Registration
Past Event
Welcome! To join the event, please register below.
About Event

Join Feihu Tang from the Moonshot AI team and Zain Hasan from Together AI for a technical session on how the Moonshot team post-trained Kimi K2 Thinking for agentic tool calling interleaved with its reasoning traces — and how to actually run it on Together.

Most models “think” and then make one tool call, then think again. This forces you to build complex orchestration layers with many API calls for each task. For real-world tasks like deep research, multi-step planning, or anything requiring multiple sequential tool calls, this architecture becomes brittle fast.

Kimi K2 Thinking takes a different approach: it can make 300 tool calls while it’s thinking, significantly simplifying complex tasks. This session will focus on how this agentic setup works in practice, how it differs from earlier open-source releases from the likes of DeepSeek and Qwen, and how to harness it to power your own agentic applications on Together.

This is a learning-first webinar — ideal for folks building and scaling LLM-powered apps, tools, and agents who need to understand what’s different about this architecture and how to use it in production.

What we’ll cover

  • Why tool calling is the bottleneck for agentic apps: the limitations of single-turn workflows and why they force you into complex orchestration patterns that don’t scale.

  • Why this model is special: what “tool calling inside the thinking trace” actually means, and why K2 doing 200–300 tool calls in one run is so powerful.

  • What Moonshot learned from training Kimi K2 Thinking for heavy tool use: high-level lessons from their agentic post-training work and how it shaped K2’s behavior.

  • Agentic workflows in one call: examples of tasks that used to need multi-step orchestration (research, planning, tool fan-out) and how K2 can keep that inside the model loop.

  • Live demo on Together AI: a “before/after” comparison showing a regular thinking model vs. Kimi K2 Thinking with interleaved tool calling on the same problem, and how to run it in production on Together.

Register to join the webinar live, ask questions, and receive a recording of the talk.

⚠️ Note: Only company (non-Gmail, etc.) email addresses will be accepted. Please provide your valid corporate email upon registration.

Avatar for Together AI Calendar
Hosted By