Technical Articles
Deep engineering write-ups from building TinyTitan: how a piece works, what was measured, and — just as often — what was tried and rejected.
You do not need any of this to use TinyTitan. If you want to install a model and run prompts, start at Getting Started; the guides and the Runtime Controls reference cover everything a user touches. These pages are for people who want the reasoning and the numbers behind the runtime: they quote exact commands, record files and per-machine measurements, and they document dead ends on purpose, because a rejected approach that is written down costs one reading instead of one re-implementation.
| Article | What it covers |
|---|---|
| ANE Prefill Research | Running attention on the Neural Engine: the 1.13–1.68× wins, the sparse-indexer fold that made Qwen 3.8 correct, and the two measured reasons it is still slower there — plus the gather graph that was sized and rejected |
A technical article earns a page when it has all three of:
- a question that a user cannot answer by reading the guides — a design trade-off, a cost model, or an approach that was considered;
- measurements, on a named machine, with the command that produced them;
-
a record committed under
benchmark/ordocs/in the source repository, so the claim can be re-run rather than believed.
Pages that explain how to use something belong in the guides instead. Pages that describe what shipped, in order, belong in the Changelog.
Note
Neural Engine behaviour is per-chip. Numbers here come from Mac15,3 (M3, 24 GB) unless the article says otherwise, and are measurements rather than ceilings — re-run the probe on your machine before treating one as a limit.
Start
Use TinyTitan
DeepSeek Harness
Reference
Engineering
Project