Released 2026-08-26 · general agent
Deployment evidence
Z.ai · 320B total · 18B per token active · 1M tokens
Best-fit scenarios
Cost-sensitive multimodal coding, document, research and tool-using agents that need a one-million-token model specification and can support either heavy host-memory offload or a multi-GPU worker.
Not recommended
- Treating the 18B active-parameter count as the memory footprint; all 320B parameters still have to be stored or offloaded
- Using the single-GPU CPU-offload route as a production baseline for a 200-person team
- Promising 1M-context concurrency before measuring both the KDA state pool and KV cache
- Leaving maximum reasoning effort enabled for every latency-sensitive request without evaluating token cost
Coding and visual software-engineering agents
Vendor-statedZ.ai positions GLM-5.3-Flash for coding and agent workflows that combine repository context, screenshots, visual debugging and tool calls. The quality claims are vendor results, not ChinaAPI reproductions.
Capability sourceLong documents, presentations and data analysis
Vendor-statedThe official release highlights long-document synthesis, presentation generation and data-analysis work with native image and video understanding.
Capability sourcePrivate multimodal research and meeting workflows
ChinaAPI inferenceThe combination of image/video input, a 1M model context and local OpenAI-compatible serving makes it a plausible fit where raw media or long records must stay inside the operator's boundary; this is a fit assessment, not a quality benchmark.
Official launch / minimum
Evidence BOfficial heterogeneous minimum
KTransformers documents a single-GPU path for the official FP8 weights on NVIDIA SM89 or SM120 hardware, with at least 350 GB of available system memory, AVX-512 CPU inference and 64 CPU inference threads. Its example uses 501,025-token context. This establishes runnable heterogeneous inference, not production latency or concurrency. Primary recipe
200-person team
Evidence BValidated worker benchmark required
For the defined 200-seat profile, start load testing with one current framework-documented worker: vLLM publishes TP4 on one 4x GB200 tray, while SGLang validates a 4x GB300 TP4/EP4 multimodal topology. Add a second independent worker only when maintenance or failure recovery is required. No official recipe maps employee count to capacity. Primary recipe
Commercial API
Evidence BOfficial serving recipe
Use at least two independently deployable workers plus admission control and separate short-, image/video- and long-context SLOs. SGLang verifies encoder disaggregation on one shared 4x GB300 node for image requests and videos up to 238,080 visual tokens; its prefill/decode path remains preview-only. Final capacity still requires measured traffic and failure tests. Primary recipe
Commercial-use check
MIT
MaaS / hosted service: No model-specific MaaS restriction identified in the MIT license published with the official weights.
Attribution: Retain the copyright and permission notice in copies or substantial portions of the weights. The weights are provided without warranty.
Known limitations and open questions
- No ChinaAPI hardware reproduction, quality benchmark or like-for-like comparison between the downloadable checkpoint and hosted route
- The KTransformers single-GPU recipe publishes no throughput or 200-user concurrency result and requires at least 350 GB of available system memory
- Its documented heterogeneous example uses 501,025 tokens rather than proving the full 1,048,576-token window under load
- vLLM currently requires a Docker build before GLM-5.3-Flash integration reaches its public repository; pin the exact serving image and framework revision
- SGLang's prefill/decode path is preview-only and has been mechanically validated with dummy weights rather than load- or accuracy-tested
- Reasoning defaults to maximum effort unless low or high is passed explicitly, which can materially change latency and billed output
- Vendor capability and efficiency claims have not been reproduced by ChinaAPI