Updated

A stack of interlocking indigo glass hexagonal blocks with thin cyan edge glow on a dark background, abstract hero artwork for GLM-5.3

Z.ai

GLM-5.3

Z.ai's open-weight, text-only model for long coding runs and security work, with 1M context.

Chat with GLM-5.3 in Super AgentView pricing

What is GLM-5.3?

GLM-5.3 is Z.ai's flagship for complex software engineering and long agent tasks, released on Z.ai's API on August 18, 2026, with open weights following on Hugging Face in late August.

Z.ai says every gain over GLM-5.2 comes from post-training on the same base model: it scores 50% higher on Z.ai's in-house Code Bench while using fewer output tokens at every effort level, and it sets open-source state-of-the-art results on Terminal Bench 3.0 and Agents' Last Exam. Z.ai also reports the best result so far on the CyberGym vulnerability-discovery benchmark. It takes text in and gives text out, with a 1M-token context window and up to 128K tokens of output, and its reasoning is always on.

Provider
Z.ai
Context window
1M tokens
Input
Text
Reasoning
Low, medium, high, max
Max output
128K tokens
Released
Aug 18, 2026

How GLM-5.3 performs

Published by Z.ai

Results Z.ai reported for GLM-5.3. Each lab runs its own setup, so compare these with other models only inside the same table.

Benchmark results Z.ai published for GLM-5.3
BenchmarkScoreSetting
Terminal-Bench 2.188.2%Claude Code harness
Terminal-Bench 3.028.3%max effort
DeepSWE v1.166.9%mini-swe-agent harness
NL2Repo58.0%
SWE-Marathon v1.142.5%max effort
CyberGym84.5%max effort
Toolathlon Verified73.0%
AutomationBench v1.0.648.2%
Agents' Last Exam (CLI)28.5%max effort
Humanity's Last Exam62.5%with tools

Source: huggingface.co

What you can do with GLM-5.3

  • Open weights, with every gain over GLM-5.2 coming from post-training

  • 50% better than GLM-5.2 on Z.ai's in-house Code Bench

  • Fewer output tokens than GLM-5.2 at every effort level

  • 1M-token context window and 128K-token output

  • Function calling and structured JSON output

How to prompt GLM-5.3

  • Describe any visual reference in words, since GLM-5.3 can't read images.
  • Give it a long, clearly ordered task; it is built for long runs.
  • Z.ai recommends max effort for complex coding. Super Agent starts it at low.

Where GLM-5.3 fits best

  • Long, text-based agent runs

  • Structured planning and outlines

  • Working through long documents pasted as text

  • Coding-style, step-by-step tasks

How GLM-5.3 compares

GLM-5.3 compared with other chat models on Flashloop
ModelLMArena textLMArena codingContext windowReasoning
GLM-5.31481#261521#331M tokensLow, medium, high, max
GPT-6.1 Sol1480#271537#141.05M tokensLow to max
Claude Opus 4.81482#251533#181M tokensOff to max
Gemini 3.6 Flash1483#221518#411M tokensLow, medium, high
Claude Sonnet 5.51478#291528#261M tokensLow to max

Good to know before you start

  • Text input only, so it can't look at images; the model picker tags it Text only
  • Z.ai documents low, high and max effort; the picker's medium setting isn't in Z.ai's docs
  • Its replies are text too, so images and video come from Flashloop's generation models

Start chatting with GLM-5.3

Chat with GLM-5.3 in Super Agent

  1. Select GLM-5.3

    Open Super Agent on Flashloop and choose GLM-5.3 in the model picker.

  2. Enter your prompt

    Ask for what you need: an answer, a plan, or the images and videos you want the agent to make.

  3. Send and keep going

    Send the message. GLM-5.3 answers in Super Agent and runs any generations you asked for.

Questions about GLM-5.3

Can GLM-5.3 read images?

No. It accepts text only, and Super Agent marks it Text only in the model picker. Pick a multimodal model if you need to attach images.

What is GLM-5.3 good at?

Complex software engineering, long agent tasks and security research. Z.ai reports open-source state-of-the-art results on Terminal Bench 3.0 and Agents' Last Exam.

Can I turn off reasoning on GLM-5.3?

No. Z.ai removed the option to disable reasoning. In Super Agent you can choose low, medium, high or max, and it starts at low.

Is GLM-5.3 open-weight?

Yes. Z.ai publishes the weights on Hugging Face under its own GLM-5.3 License.

More chat models on Flashloop

Browse all models