跪拜 Guibai
← All articles
Backend · Frontend · GitHub

Zhipu Ships GLM-5.3-Flash: 300B MoE Model at 1/40 the Price of Opus 4.8

By 程序员_小雨 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Top-tier model capability at 1/40 the price of Opus 4.8 removes the token-counting constraint for solo devs and small teams. Native screenshot-to-code vision eliminates the design-draft ping-pong that slows frontend work.

Summary

The anonymous Ox-Alpha model that stirred overseas forums is Zhipu’s GLM-5.3-Flash, now fully available. It packs 300B total parameters under a Mixture-of-Experts architecture and surpasses the prior flagship GLM-5.2 across complex engineering tasks, terminal operations, and agent workloads — matching Opus 4.8 while activating far fewer parameters.

The pricing resets expectations: the regular rate is 1/10 of the full GLM-5.3, and a limited-time discount drops it to 1/20, making it about 1/40 the cost of Opus 4.8. Solo developers and students can run batch tests, large-repo analysis, and automated agent jobs without rationing tokens. Existing Coding Plan subscribers get their quota tripled when using this model.

GLM-5.3-Flash is the first natively multimodal model in the GLM line. It reads screenshots, rendered designs, and finished page images directly, producing higher-fidelity frontend code and cutting style-adjustment time by more than half. Office document generation — PPTX, PDF, DOCX, XLSX — ships as finished files rather than raw text needing manual formatting.

Takeaways
300B total parameters with a MoE architecture; active parameters are low but performance exceeds the previous flagship GLM-5.2 and matches Opus 4.8.
Regular pricing is 1/10 of GLM-5.3; a limited-time discount brings it to 1/20, roughly 1/40 of Opus 4.8.
Coding Plan subscribers get 3× quota when using GLM-5.3-Flash.
First natively multimodal GLM model — it reads screenshots and rendered images directly, producing layout-aware frontend code.
Supports direct output of PPTX, PDF, DOCX, and XLSX files, not just raw text.
A free 7-day trial card is available for unsubscribed users, limited to 10,000 per day.
Conclusions

A Flash-tier model beating the previous flagship suggests Zhipu’s distillation or training efficiency has jumped, not just its parameter budget.

Pricing at 1/40 of Opus 4.8 turns large-model access from a metered expense into a flat, negligible cost — this changes how freely developers can integrate LLMs into CI, batch processing, and agent loops.

Native multimodal input that reads screenshots sidesteps the brittle text-description bottleneck that makes current AI coding assistants stumble on visual layout tasks.

Concepts & terms
MoE (Mixture of Experts)
A transformer architecture where only a subset of specialized sub-models (experts) activate per token, keeping inference cost low while scaling total parameters into the hundreds of billions.
Native multimodal model
A model trained end-to-end on multiple modalities — text, images, possibly audio — so visual understanding is baked into the core weights rather than bolted on via a separate vision encoder post-training.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗