GLM-5.3-Flash Architecture Notes

Now we know: The popular Ox Alpha LLM was GLM-5.3-Flash…

Compared to GLM-5.2, this new GLM-5.3-Flash model uses:

a Kimi Linear-style 3:1 (super*) hybrid attention pattern with 34 Kimi Delta Attention layers (KDA) and 11 Multi-head Latent Attention (MLA) / DeepSeek Sparse Attention (DSA) layers; #

a scaled-down GLM-5.2-style sparse MoE backbone, going from 744B-A40B to 320B-A18B; #

a DeepSeek V4-style mHC residual path with four parallel streams;

plus a native vision encoder (not shown).

I called it a “super hybrid” above because both KDA and MLA/DSA are “efficient” components. E.g., Kimi only uses KDA + full attention MLA, DeepSeek V3.2 uses DSA + full attention MLA.

PS: I’m sorry for the excessive tech jargon. Explainers on all these components (MLA, DSA, KDA, mhC, etc.) in my LLM Architecture Gallery. PPS: Haha, maybe justification for getting that pricey Mac Studio M5 Ultra 256 GB / 512 GB to run this locally…

Source: website version of my Substack note.

Read Next #

How Claude's Text Watermarking Works Short illustration of how Claude's text watermarking is supposed to work based on Anthropic's released materials.

Build a Reasoning Model From Scratch Is Now on Amazon Short note on the Amazon availability of Build a Reasoning Model From Scratch and a warning about counterfeit black-and-white copies sold through Amazon In

Muse Glimmer 30B Architecture Notes Short architecture note on Meta Muse Glimmer 30B, including gated local and global GQA, KV-cache efficiency, and release-time benchmark comparisons.