Kimi K3 Quantization Lab
An early technical laboratory exploring how far Kimi K3's open weights can be compressed for modest hardware while preserving as much capability and usability as possible.
- Kimi K3
- Mixture of Experts
- Quantization
- MXFP4
Hi, my name is
I build pragmatic software; Go backends, React frontends, and full-stack products across web, mobile, and desktop. Currently based in Hermosillo, Sonora, Mexico.
Chat with Rony
Rony — Victor Hugo Vargas Servin's dog AI assistant
Note: Rony runs on a very small model using only 2 CPU cores, so replies may take a little longer.
An early technical laboratory exploring how far Kimi K3's open weights can be compressed for modest hardware while preserving as much capability and usability as possible.

A terminal-first AI agent harness that brings local and cloud models, real software tools, human approvals, and persistent sessions into one focused workspace.

Interactive demo that lets visitors pick which Resume sections, jobs, and skills to include in a downloadable PDF — generated entirely in the browser from a structured expertise collection.


Chat widget on this very site that uses RAG over Markdown project descriptions to answer questions about the work shown here. React frontend with SSE streaming; Astro backend as a proxy.

Go HTTP service that wraps an LLM agent and streams responses to web clients over server-sent events. Configurable persona (Rony), persistent storage, MCP-style tool calling, and context-window management with on-demand compaction.

Core LLM agent library in Go: provider-agnostic client interface, agent loop with iteration control, retrieval-augmented memory backed by ChromaDB, tool registry with a filesystem sandbox.