GLM-5.3-Flash Debuts as Zhipu AI's First Natively Multimodal Model, 320B Parameters with 18B Active

What was released

Zhipu AI (zai-org) has released GLM-5.3-Flash, described as the first natively multimodal model in its GLM-5 series. The model uses a mixture-of-experts (MoE) architecture with 320 billion total parameters but only 18 billion active parameters per inference pass. Unsloth has published quantized GGUF versions of the model (unsloth/GLM-5.3-Flash-GGUF) for local deployment, alongside a companion guide for running it, claiming their Dynamic 3.0 quantization method preserves accuracy better than other available quants.

Architecture changes

According to Zhipu AI, GLM-5.3-Flash starts from a newly trained base model rather than a fine-tune of a prior checkpoint. The company introduced a hybrid attention architecture combining sparse and linear attention for the first time in the GLM series, which it says sharply reduces long-context serving costs while preserving long-context accuracy. The model also adopts Manifold-Constrained Hyper-Connections (mHC), a technique Zhipu says improves scaling efficiency. Pre-training used a 30-trillion-token multimodal corpus.

Performance claims

Zhipu AI claims GLM-5.3-Flash outperforms its predecessor, GLM-5.2, across benchmarks and real-world workloads while running at one-tenth the price. The company also states the model approaches Claude Opus 4.8 on coding and agentic benchmarks. These comparisons come from Zhipu's own technical report and have not been independently verified.

The report cites evaluation on several benchmarks — HLE (with tools), NL2Repo, DeepSWE, Terminal-Bench 2.1, Agent's Last Exam (Toolathlon Verified), AutomationBench v1.0.6, GDPval-AA v2, and BabyVision — but the source material does not include the numeric scores themselves, only evaluation methodology (e.g., NL2Repo tested under 1M-token context with temperature=1.0; HLE evaluated at up to 300,000-token context using GPT-5.6-luna as judge model).

Availability and cost

GLM-5.3-Flash is accessible via the Z.ai API Platform; specific per-token pricing has not been disclosed in available materials. Quantized GGUF builds from Unsloth are available on Hugging Face for local inference, though as of publication no inference provider had deployed the model for hosted API access.

What this means

GLM-5.3-Flash extends the trend of MoE models decoupling total parameter count from active compute cost — at 18B active out of 320B total, it's designed to deliver large-model capability at small-model serving cost. The hybrid sparse-linear attention approach is a notable architectural bet on cheaper long-context handling, an area where dense-attention models incur steep serving costs at scale.

The claim of approaching Claude Opus 4.8 on coding and agentic tasks is significant if verified independently, but it comes solely from Zhipu's own technical report without published benchmark numbers in the source material reviewed here. Buyers and developers should treat the Opus-4.8 comparison and the "one-tenth the price" claim as unverified until third-party benchmarks or confirmed API pricing are available. The rapid availability of Unsloth GGUF quantizations, however, signals that the open-weight community expects strong developer interest in running this model locally.