Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context
1M-token context, 18B active parameters, MIT-licensed: Z.ai's GLM-5.3-Flash is a natively multimodal MoE that cuts attention compute 3× and KV cache 4.4× versus its predecessor.
Why it matters
A significant open-weight multimodal model release with aggressive efficiency gains (MoE, sparse attention, NoPE) and a 1M-token context window — practitioners can now deploy native multimodal reasoning at scale without proprietary APIs, and the architecture innovations (hybrid KDA + NoPE sparse MLA) are worth studying for cost-optimized deployments.
The key facts
10 to know320B total parameters / 18B active (MoE)
1,048,576-token context window
MIT-licensed weights on Hugging Face
API pricing: $0.15/M input, $0.50/M output
Terminal-Bench 2.1 score: 84.3
DeepSWE v1.1 score: 63.4
Attention compute reduction: ~3× vs GLM-5.3
KV cache reduction: 4.4× vs GLM-5.3
Architecture: hybrid KDA linear + NoPE sparse MLA attention
Natively multimodal (first in GLM-5 series)
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series — a 320B-total / 18B-active MoE with a 1,048,576-token context window, MIT-licensed weights on Hugging Face, and API pricing at $0.15/M input and $0.50/M output. It scores 84.3 on Terminal-Bench 2.1 and 63.4 on…