Articles for people who
ship the stack
Original rewrites on React, Node.js, TypeScript and AI — practical notes from the same engineering practice behind our operator software. Article bodies are in English.
Tagged: Local AI
Serving Qwen3.8-27B on One RTX 3090 With a Patched vLLM and DFlash2
How a pinned vLLM 0.28.0 fork with requantized embeddings and DFlash2 speculation serves a 27B hybrid model on a 24 GB card, and where long context slows it down.
15 min readRecurrent Depth Explained: Looped Transformers and Hidden Reasoning
How recurrent-depth (looped) transformers reason in hidden state, why labs want them, what eight years of papers show, and what it means for local AI and oversight.
12 min readServing Qwen3.8-Flash-Next on RTX 3090s with llama.cpp Tensor Offloading
Why an 88GB, 180B-parameter mixture of experts model like Qwen3.8-Flash-Next fits RTX 3090s, and how llama.cpp -ot, -ncmoe and mmap split VRAM, RAM, NVMe.
12 min readWhy a 17 GB Model Download Doesn't Mean 17 GB of Memory to Run It
Learn how active parameters, total parameters, KV cache growth and reasoning effort decide the real memory and compute cost of running a language model locally.
12 min readBuilding a Local Angry Birds Clone With Qwen3.8-27B and Pi
A fully local AI coding setup with LM Studio and the Pi agent: Qwen3.8-27B builds a playable Angry Birds level, no cloud API, with every install step shown.
9 min readWhy AI Access, Not Capability, Is Your Real Dependency Risk
This article examines recent export-control incidents with Claude and GPT-5.6 to argue that access to AI models is a volatile variable independent of raw capability.
11 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.