Tutorials Self-Host Kimi K3 Open Weights With vLLM
Run Moonshot's 2.8T open-weight model on your own GPUs with vLLM and MXFP4.
How-to content for builders, indie hackers, and AI engineers. Less theory, more shipped code.
Tutorials Run Moonshot's 2.8T open-weight model on your own GPUs with vLLM and MXFP4.
Tutorials Use Kimi K3 tool_choice=required + strict JSON so agents fetch data before they answer.
Tutorials Prefill Kimi K3's assistant turn to lock output shape, force clean JSON, and steer tone.
Tutorials Tune Kimi K3's low/high/max reasoning effort and stream reasoning_content to control token cost.
Tutorials Load tools just-in-time in Kimi K3's agent loop to shrink prompts and sharpen tool choice.
Tutorials Use Kimi K3's native vision and JSON schema output to turn messy PDF pages into clean, typed data.
Tutorials kimi-k3 is not a drop-in swap. Map the params right and dodge the trap that breaks tool loops.
Tutorials Point the OpenAI SDK at Moonshot's 2.8T K3, load a whole repo, and cut cost with caching.
Tutorials Drive Moonshot's open-weight coding model through a real tool-calling loop in Python.
Tutorials Build a multi-step tool-calling agent on Moonshot's open-weight Kimi K2.6 model.