Writing
On Medium, at blog.0xbakeer.com. Most articles begin with a run on my own machine, and the ones I like best begin with a run that went wrong.
TandemLLM
How the engine drafts and verifies, how StairCut sizes the draft tree each round, and what the paper measured. The project
DeepSeek-V4.1-Flash on one DGX Spark
A series on expert offloading, in order. Which experts stay in memory, why held-out loss did not see the damage, and what the model does when a task needs an expert that is not there. The project
- Part 1Which 139 of 384 experts, and who decides
- Part 2Held-out loss said the broken model was better
- Part 3Frequency is not contribution
- Part 4One vote each is not one outcome each
- Part 5Ten profiles, one box, and what each one keeps
- Part 6Long answers find the experts I left out
- Part 7A 510 GB model on a 128 GB box, measured
arbiter
Why I built a local Jev on a Sunday morning, and how arbiter answers typed questions from an encoder that sits next to a language model. The project
More articles
Seven ways to serve Qwen-Image-2.1, and a bug hunt in vLLM's attention path.