GLM-5.3

Checkpoint identity

Official identifier

zai-org/GLM-5.3Z.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Architecture

Text mixture-of-experts model; GlmMoeDsaForCausalLM checkpoint architecture with sparse attentionZ.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Total parameters

Not verified for this recordPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Z.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Active parameters

Not verified for this recordActive parameters are a compute-related count, not the model's full stored weight count.Z.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Modalities

text-input, text-outputZ.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Context

1,048,576 tokensNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Z.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Release date

Exact release date not independently established in the reviewed artifactsZ.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Download access

Public checkpoint referenceZ.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

License

GLM-5.3 License (custom)This version has its own license rather than an unmodified MIT license. Notices and applicable-law duties apply. A licensee or affiliate operating Model as a Service must pass Z.ai's security review before commercial use if aggregate licensee-and-affiliate revenue exceeds US$10 billion over any consecutive 12 months. Read the definition and exact version-specific terms before commercial deployment.Z.ai GLM-5.3 official model card (external source) · GLM-5.3 checkpoint configuration (external source) · GLM-5.3 version-specific license (external source) · SGLang model-specific GLM-5.3 cookbook (external source) · vLLM model-specific GLM-5.3 recipe (external source)

Theoretical weight-memory footprint

A weight estimate is unavailable without a verified parameter count.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

The main checkpoint configuration declares blockwise FP8 (E4M3, 128 by 128 weight blocks), while the general dtype is BF16 and many modules are excluded from conversion. The publisher also lists a distinct BF16 artifact. Do not treat every tensor as FP8 or substitute a different artifact without evaluation.

Serving considerations

• Exact artifact identity matters: GLM-5.3 is a separately licensed post-training release that its card says shares the GLM-5.2 base. Its total and activated parameter counts are left unknown here because this card does not document those counts unambiguously.
• The checkpoint declares 1,048,576 maximum positions. This is a configuration boundary; long-context task quality, cache allocation and supported serving limits still need workload-specific evaluation.
• The official card points to model-specific vLLM and SGLang recipes. These are publisher-documented deployment paths, not Aguasto tests. Pin model revision, engine build, attention kernels and quantization support before measuring memory or throughput.
• The card defaults reasoning_effort to max and clear_thinking to false. Request formatting and reasoning settings can change retained context and output-token demand; record them when comparing latency or cost.
• No Aguasto hosting, GPU fit, throughput, API availability or operational benchmark is represented. The profile documents a candidate for technical evaluation.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.