GLM-5.3-Flash

321B total, 18B active, hybrid KDA and sparse MLA, 288 experts top-8. No engine burns it yet.
Modified

2026-09-05

property value
parameters ~321B total, 18B active per token
layers 45 text, plus a 24-layer vision encoder
attention KDA linear on ~34 layers, NoPE sparse MLA with a top-2048 DSA indexer on every 4th layer
MoE 288 routed experts, top-8 per token
context 1,048,576 tokens
speculation native MTP draft layer; DFlash2 external drafter also published

Reusable state has two parts. Sparse MLA KV grows with sequence length. KDA recurrent state is fixed size per sequence and cannot be reconstructed from KV.

Weight formats

format size fits 1 node (123.73 GiB) fits 2
FP8 native ~306 GiB no no
NVFP4, 4-bit experts ~182 GiB no yes
EXL3 4bpw ~164 GiB no yes

No format fits one node, so any engine for this fuel is split across at least two.

On disk here: Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw (164 GB) and incoai/GLM-5.3-Flash-DFlash2 (2.2 GB drafter).

What it demands

18B active parameters at ~4.25 bits is about 9.6 GB read per token if every active weight is re-read. Against one node at 273 GB/s that is a ~28 tok/s single-stream ceiling. Two nodes working the same token halve the per-node read and roughly double it.

Only 8 of 288 experts fire per token, so tokens that share experts can amortize the read. That is where aggregate throughput above the single-stream ceiling comes from.

Engines

None yet.