GLM-5.3-Flash: Multimodal Architecture, License and Benchmarks

A flash model with its own architecture, not a distillation

GLM-5.3-Flash carries 320 billion total parameters with 18 billion active per token, a mixture-of-experts model with 45 layers and 8 of 288 experts firing on a given token. Z.ai describes it as trained from a new base model rather than distilled from the flagship GLM-5.3, with architecture and training recipe redesigned around efficiency from the start. [1]

The model pairs linear attention, which captures nearby dependencies through state modeling, with sparse attention, which retrieves relevant context from farther back through a lightweight indexer. Z.ai calls the combination new for the GLM series. A component it names IndexPool compresses four indexer key vectors into one through weighted pooling, aimed at keeping that sparse retrieval cheap at a million-token context. [1][2]

Z.ai's own comparison puts GLM-5.3-Flash's attention compute at roughly a third of GLM-5.3's per head per layer, with a key-value cache about a quarter the size. Those are the vendor's numbers, measured against its own prior model rather than an independent benchmark of inference cost. [1]

A week running anonymously before the name was attached

Before naming the model, Z.ai ran it anonymously as "ox-alpha" on OpenCode and OpenRouter to gather feedback under ordinary use rather than a controlled benchmark. Z.ai's own account says the model became the most popular of the week under that alias, all of the traffic served on Chinese AI chips, and names the release as the reveal. [1]

SiliconANGLE's account differs at the edges: it describes OpenRouter offering the free hosted alias without disclosing a developer, and says outside observers, not Z.ai, first pieced together the model's origin from its behavior before Z.ai confirmed it. Either way, running a candidate model under an unbranded codename on public inference platforms ahead of a named release is a pattern this timeline has already recorded once, with GLM-5.3's own predecessor debuting through a similar anonymous test. [3]

MIT licensed at release, unlike the flagship it undercuts on price

GLM-5.3-Flash's weights are on Hugging Face under an MIT license from the day of release. Z.ai's own post confirms the weights are public and names supported inference frameworks, including SGLang, vLLM and TokenSpeed, without stating the license itself; MarkTechPost's reporting is the source for the MIT terms specifically. [1][2]

Atlas interpretation: The flagship GLM-5.3 shipped API-only and under a license Z.ai wrote for the occasion, with open weights promised only after a safety-hardening pass of about two weeks. Flash skipped that gap entirely: MIT terms and public weights on the same day the model was named. A smaller, cheaper model reaching open distribution before the larger one it was built alongside is the kind of sequencing worth noting on its own, separate from what either model can do. [1]

Vendor-reported gains over GLM-5.2, at a lower price

Z.ai reports GLM-5.3-Flash scoring 63.4 against GLM-5.2's 46.2 on DeepSWE v1.1, and 48.8 against 26.2 on AutomationBench v1.0.6, both benchmarks it groups under coding and agentic tasks. It also reports approaching Claude Opus 4.8 on several of the same tests, for instance 84.3 to Opus 4.8's 85.0 on Terminal-Bench 2.1. These are Z.ai's own published figures, not an independently run comparison, and the company sets the evaluation conditions, including which harness and context length each benchmark used. [1]

On price, Z.ai lists GLM-5.3-Flash at $0.15 per million input tokens and $0.50 per million output tokens, and separately claims it reaches a given score on the Artificial Analysis Intelligence Index at roughly a tenth of the cost that level of performance previously required. MarkTechPost's independent writeup and SiliconANGLE's both describe the model as roughly ten times cheaper to run than Z.ai's prior flagship, tracking the vendor's own framing rather than a separate cost audit. [1][2][3]