GaⅤin.com by Gavin W.
Colibrì vs. FreeToken: Can Consumer Hardware Really Run Frontier MoE Models?
An empirical comparison of Colibri vs FreeToken for edge MoE serving, breaking down RAM limits, NVMe Direct DMA, the WSL2 I/O tax, and real tokens/sec.
AI
Does Windows Dev Drive Actually Speed Up Local LLMs? A Deep Dive into ReFS, Defender, and NVMe Paging
Empirical benchmarks comparing Windows Dev Drive (ReFS) vs NTFS for local LLMs across cold TTFT, inference jitter, Colibri MoE streaming, and n-gram paging.
Getting 10 tok/s on Qwen 3.8 Flash Next: Why N-Gram Disk Paging Works Where MoE Fails
How I got Qwen 3.8 Flash Next running at 10 tokens/sec on an RTX 5060 Ti, and why paging n-gram tables from NVMe avoids the brutal disk bandwidth bottlenecks of MoE expert streaming.
Exposing Local LLMs to the Internet with LM Studio, LiteLLM, and Cloudflare Tunnels
How to securely publish local LM Studio models to the internet using a Dockerized LiteLLM proxy and Cloudflare Tunnels.
Visualizing the Training of a Custom LLM
A visual deep-dive into training a custom transformer LLM from scratch with PyTorch, analyzing loss curves, overfitting, and hyperparameter tuning.
Running Clawbot on Local LLMs and Cloudflare Tunnels
A hands-on setup guide for running Clawbot against local LM Studio models and securely exposing the web interface using Cloudflare Tunnels.
The Battle of the Agentic IDEs: Amazon Kiro vs. Google Antigravity
A "script kiddie" explores the battle between Amazon Kiro and Google Antigravity, comparing the rigid efficiency of spec-driven coding against the autonomous magic of agent-first orchestration. Find out which agentic IDE fits your "vibe coding" style and which Star Trek captain matches your workflow.