Latest Posts
Thoughts on engineering, testing, and building reliable systems.
Qwen3.8-Flash-Next: the Qwen4 architecture, two weeks early
Qwen open-weighted Qwen3.8-Flash-Next on August 26: a 125B MoE with 6B active parameters that previews the Qwen4 architecture. It beats last season's Claude flagship on the agentic benchmarks and runs on my Mac Studio at 37 tokens/s.
DwarfStarLauncher: a GUI for DS4, built with the model it serves
I wrote a native SwiftUI launcher for ds4, using DeepSeek V4 Flash served by ds4 as my coding assistant. No cloud APIs for inference.
DeepSeek V4 Flash 0731: 284B parameters on a desktop machine
DeepSeek V4 Flash landed in late April; the 0731 checkpoint was shipped on July 31. The release comes with native MXFP4 weights, and it's a good model for its size that runs on 'consumer' level hardware — with the 2026 RAM price crisis as the caveat.
Running 284B AI Models on Your Own Hardware: The DS4 Project and Local Inference
How antirez's DwarfStar engine, DeepSeek V4 Flash's MoE architecture, and aggressive quantization are making capable AI accessible outside the cloud — and what it means for developers who care about privacy, cost, and control.
Property-Based Testing: Write Fewer Tests, Find More Bugs
Property-based testing replaces hundreds of hand-written test cases with a few lines of code and finds more bugs. It's also a natural fit for the age of AI agents.