GLM-5.3-Flash is Z.ai (Zhipu) ’s efficient multimodal model in the GLM-5 family: a sparse MoE with 320B total / 18B active parameters, native vision in the coding loop, and a 1M-token context window. Searches for glm 5.3 flash, glm flash 5.3, and opencode glm 5.3 flash usually want model IDs, price, and how Flash differs from FlashX.
Model snapshot
- API model codes:
glm-5.3-flashandglm-5.3-flashx - Architecture: hybrid sparse + linear attention MoE (320B / 18B active)
- Context: 1M tokens; max output up to 128K tokens (per Z.AI docs)
- Modalities: text, image, video, and file inputs → text output
- Open weights: Flash is described as open-source/open-weight in Z.AI’s model overview; weights also appear on Hugging Face under
zai-org/GLM-5.3-Flash
Z.AI positions Flash as stronger than GLM-5.2 at a much lower serving cost, with attention/KV-cache reductions versus GLM-5.3 called out in the docs (approximately 3.01× / 4.44× less attention compute and KV cache in their comparison language).
Official API pricing (USD / 1M tokens)
From Z.AI’s pricing page:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 |
| GLM-5.3-FlashX | $0.37 | $0.075 | $1.25 |
| GLM-5.3 (ref.) | $1.40 | $0.26 | $4.40 |
Cached-input storage is listed as limited-time free. FlashX is the higher-speed tier—Z.AI markets ~200 tokens/s inference for FlashX for snappier interactive coding.
How to call it
Use the Chat Completions-style API documented by Z.AI (https://api.z.ai/api/paas/v4/chat/completions in third-party and partner docs; confirm the current base URL in Z.AI’s quick start). Recommended sampling from the model page: temperature: 1, top_p: 0.95, reasoning_effort: max, with thinking enabled. For images, add image_url content blocks.
{
"model": "glm-5.3-flash",
"messages": [{"role": "user", "content": "Hello"}]
}
GLM Coding Plan notes
Z.AI says GLM-5.3-Flash is fully available on the GLM Coding Plan with 3× quota versus GLM-5.3 on that plan. FlashX is documented as not yet on the Coding Plan. Off-peak and weekend calls can consume 50% points under the plan’s points system—check the live plan pages for Personal vs Team details.
OpenCode and similar coding agents that speak OpenAI-compatible endpoints can typically point at Z.AI with the glm-5.3-flash model id once you supply an API key.
Capabilities called out by Z.AI
Beyond raw price, the model page emphasizes:
- Visual coding loop — observe UIs and rendered results, then iterate
- Tool use / agents — break down goals, call tools, and review outputs toward Office-style deliverables (PPTX, PDF, DOCX, XLSX)
- Serving efficiency — marketed as suitable for large-scale serving, including deployments on Chinese AI chips in Z.AI’s narrative
Those claims are vendor-facing; validate on your own coding and document workloads before locking a provider.
When to use Flash vs FlashX vs GLM-5.3
- Flash — default for cost-sensitive coding, agents, and multimodal loops
- FlashX — when latency/tokens-per-second matter more than unit price
- Full GLM-5.3 — when you need the larger non-Flash SKU and can pay ~10× Flash input rates
Bottom line
GLM-5.3-Flash is Z.ai’s aggressive price/performance multimodal coding model: open weights, 1M context, and API pricing that undercuts the full GLM-5.3 SKU by a wide margin. Start with glm-5.3-flash; switch to glm-5.3-flashx only if you need the speed tier and accept the higher token rates.
Sources
- Z.AI docs: GLM-5.3-Flash/FlashX
- Z.AI docs: Pricing
- Z.ai product site
- zai-org/GLM-5.3-Flash on Hugging Face
Sources verified Oct 17, 2026 (Pacific Time)
