GLM-5.3 Flash

Z.AI's fast reasoning and tool-use model: near-frontier agentic scores at a fraction of frontier prices, with a million tokens of context.

GLM-5.3 Flash is the cost-optimised cut of Z.AI's GLM-5.3 line — a 320B mixture-of-experts that activates 18B parameters per token, built from a newly trained base rather than a distillation. The result is a model that reasons and uses tools like something much more expensive.

Its distinguishing strength is multi-step work: terminal sessions, tool pipelines, automation flows. On automation benchmarks it leads models that cost ten times as much — which is exactly the shape of workload that benefits most from post-training on your own environment.

z-ai/glm-5.3-flash

What’s it good for

Agents and tool use

Near-frontier scores on terminal and tool-orchestration benchmarks. The strongest choice per dollar for multi-step agent loops.

Automation workflows

Leads its benchmark set on multi-step automation — the repetitive, structured work where post-training pays back fastest.

High-volume reasoning

Reasoning-model quality at $0.25 per million output tokens. Built for workloads where unit economics decide the architecture.

Benchmarks

Figures as reported by Z.AI, October 2026.

Terminal-Bench 2.184.3Opus 4.8 scores 85.0

Toolathlon Verified78.4

DeepSWE v1.163.4pass@1

AutomationBench48.8leads Opus 4.8 at 41.0

Humanity's Last Exam55.3with tools

Pricing & Specs

Latency is a choice, not a property of the model — pick the service tier your workload needs.

Input / 1M: $0.075
Cached input / 1M: $0.015
Output / 1M: $0.25
Context: 1M

Weights: Open — zai-org/GLM-5.3-Flash
Released: August 2026
Architecture: 320B MoE, 18B active
Quantization on Gradiated: FP8
Max output: 131K tokens
Streaming: Yes

Getting Started

Point your existing client at our base URL and name the model. Your traffic starts accruing as traces — the raw material post-training is built from.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.gradiated.com",
    api_key=os.environ["GRADIATED_API_KEY"],
)

response = client.chat.completions.create(
    model="z-ai/glm-5.3-flash",
    messages=[{"role": "user", "content": "Summarize this document."}],
)

Read the quickstart

Hardware

What GLM-5.3 Flash runs on — our serving fleet first, then what the open weights support anywhere else.

NVIDIA B300: What Gradiated serves on — tensor-parallel at FP8
NVIDIA H100 / H200: Self-host in BF16 or FP8
Inference engines: vLLM, SGLang, KTransformers, Transformers