Tuan Pham
SYS // OK
MENU

Tuan Pham — Director of Software Engineering

00 // About this place

Welcome to my corner of the web, a place for the things I build, the ideas I’m exploring, the lessons I’ve learned, the curiosities I keep chasing, and the art of making complex things work.

Value loaded // ready
tpham@ubuntu: ~/portfoliobash
01 / Writing

Log Records

Architecture notes, leadership thinking, and things I learned the hard way.

3 / 12 records
LOG-001DRAFT

Strata on a Mac Studio M5 Ultra (96 GB)

Strata first claimed Qwen3.8-Flash-Next, a 125B model, could run on a machine with just 12 GB of VRAM. My new Mac Studio M5 Ultra with 96 GB just arrived, so I turned it into a headless local AI inference machine.

Tag
/llm /local-ai
Date
Read
9 min read
LOG-002LIVE

Dense vs. MoE Models Explained

A 30B model with 3B active parameters still needs room for its full weights. Here is how dense and mixture-of-experts models differ and what total parameters, routing, memory, and compute mean for local AI.

Tag
/llm /local-ai
Date
Read
10 min read
LOG-003LIVE

Prefill vs. Decode

Prefill, or prompt processing (PP), reads the input. Decode, or token generation (TG), writes the output. GPU compute matters for ingesting repository context; memory bandwidth helps explain generation speed.

Tag
local-ai
Date
Read
6 min read
Explore all 12 writing records →
02 / Lab

Lab Records

Things I build, break, and benchmark outside of work. Ongoing experimentation, not a polished portfolio.

3 / 13 records
LAB-013TESTING

Qwen3.8 Flash-Next 125B: Mac Studio M5 Ultra Coding Benchmark

A single-shot coding benchmark for ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF at IQ3_S, served by Strata on a Mac Studio M5 Ultra with 96 GB of unified memory at 256K context, and driven through two harnesses: the VS Code Copilot harness and OpenCode. One prompt builds an elaborate voxel-art pagoda garden as a single HTML file, with the full machine setup documented in the Strata write-up.

Stack
Qwen3.8 Flash-Next 125B, Strata, VS Code Copilot harness, OpenCode, Mac Studio M5 Ultra
Date
2026-10-24
Source
Local inference benchmark
Launch tool
LAB-012TESTING

Qwen3.8 27B — RTX 5080 Coding Benchmark

A local coding benchmark for ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF at IQ3_S-MTP on an NVIDIA RTX 5080 with 16 GB of VRAM. Three single-file web builds test responsive styling, animated canvas generation, and interactive UI logic while recording the exact inference configuration and decode performance.

Stack
Qwen3.8 27B, llama.cpp, RTX 5080
Date
2026-10-06
Source
Local inference benchmark
Launch tool
LAB-001ACTIVE

LLM Inference Capacity Planner

An interactive capacity-planning tool for local LLM inference. It estimates model weight storage, KV cache growth, runtime VRAM, minimum host memory, and GPU count from model size, quantization, context length, cache precision, batch size, and the target memory architecture. The breakdown makes the assumptions visible so a hardware plan can be adjusted before downloading a model or buying a GPU.

Stack
Next.js, TypeScript, LLM memory modeling
Date
2026-10-04
Source
Interactive lab tool
Launch tool
Explore all 13 lab records →
03 / Uses

Systems I build with

Three purpose-built systems, backed by a compact lab and local AI stack.

Full manifest →
SYS-01Primary workstation

Primary workstation

CPU
Intel Core i9-14900K
Memory
64 GB RAM - 6400MT/s
SYS-02Gaming system

Gaming Rig

CPU
AMD Ryzen 7 9800X3D
Memory
96 GB DDR5 RAM - 6400MT/s
SYS-03AI coding system

AI Agentic Engineering Node

System
Mac Studio
Chip
Apple M5 Ultra
Homelab12 nodes // 6 systems
AI runtimellama.cpp / Open WebUI / vLLM