About

the dimsum is a weekly publication on AI engineering intelligence from Chinese labs.

The frontier conversation tends to process these labs through press releases and leaderboard screenshots. We go the other way: model cards, technical reports, arXiv papers, repository contents, rate cards, and license files — read directly, every week, and synthesised for engineers who act on the details.

What we cover

• Model architecture — attention and memory-system design, MoE routing, multi-token prediction, long context.
• Inference and serving — kernels, frameworks, speculative decoding, pricing surfaces that matter for agent loops.
• Open weights — licenses read from the artifact, checkpoints inspected, serving-stack support checked.
• Independent evidence — benchmarks cross-checked, vendor numbers always labelled as vendor numbers.

How we work

Every issue follows the same pipeline: research → verification → editorial synthesis → human review → publication. Every checkable fact traces to a ledger row with its source and evidence tier before it enters a draft — see Methodology for the full description.

Issues publish weekly, on Sundays. The feed is available as RSS.

Corrections

We do not silently edit material factual errors. Corrections ship in a dated Corrections section at the bottom of the affected issue, with the primary source each correction was verified against.