GLM-5.3-Flash: 18B active out of 320B, and a tenth-price claim that needs reading

The design choice worth reading here is the attention stack, not the headline parameter count. Z.ai says this is the first model in the GLM series to combine sparse and linear attention, and that the purpose of the combination is long-context serving cost while keeping precise long-context behaviour ^[5]. Around it sit Manifold-Constrained Hyper-Connections for scaling efficiency ^[6], a newly trained base model ^[4], and a 30T-token multimodal pre-training corpus ^[7]. On the compute side the arithmetic is plain: 18B of 320B works out at 5.6% of the weights active per token, with 94.4% untouched on any given forward pass ^[1]. That ratio is what a tenth-of-the-price API listing rests on ^[3][9], and it is a statement about work per token rather than about everything else on the invoice.

The evaluation footnotes are the most useful part of the card, because they describe the conditions under which the coding claim was measured. NL2Repo ran at 1M context with 64k new tokens ^[11]. DeepSWE ran in the mini-swe-agent harness at 400K context with a six-hour timeout ^[13]. Terminal-Bench 2.1 ran inside Claude Code 2.1.207 with 65,536-token generations and the same six-hour ceiling ^[14]. HLE with tools used a 300,000-token window plus a context management strategy, with GPT-5.6-luna at medium acting as judge ^[10].

Not all of those numbers are self-administered. GDPval-AA v2 was scored by Artificial Analysis ^[15], and Toolathlon Verified went through the official evaluation service, reported as pass@1 averaged over three runs ^[16]. The rest sit inside the vendor's own harness choices, which is ordinary practice, and also the reason the judge model and the context strategy matter as much as the resulting figure.

One footnote is a deployment note wearing benchmark clothes. NL2Repo judging combined rule-based and LLM-based checks specifically to catch unauthorized pip or curl operations ^[12]. A team writes that check because the model reaches for the network while it works, which tells you something about the sandbox you need before you run it against a real repository.

The scores themselves are not in the card text supplied here, only the settings that produced them ^[18]. That leaves three things in the pitch: fewer FLOPs per token, cheaper long-context serving, and parity with Claude Opus 4.8 on coding and agentic benchmarks ^[3]. The first is arithmetic, the second is measurable on rented hardware with SGLang or vLLM ^[8], and the third is currently a sentence.