▍ Local inference · agentic systems · Tartu
This is where I write about that work — what I'm testing, what breaks, and what I find. Diary-shaped, evidence-first. No conclusions I haven't run.
A pruning pipeline running on my own hardware. 50M tokens in, a checkpoint to dissect, and a live question: where does uniform expert pruning stop being free, and what's the real VRAM-vs-quality sweet spot?
Read the developing notes →Summer 2025, one GPU bought on a hunch, a €9,000 receipt and the regret that came with it — and the project that turned it all into four Blackwells. The build, in photos.
Read the origin story →I went looking for decode speed and found something else: quantized MoE inference loses determinism in proportion to how aggressively you quantize — and a leaked timestamp in my own harness was faking half the signal. A controlled walk across seven configurations.
Read the finding →