What are the strengths of this model?

SOTA open-weight agentic coding (Terminal-Bench 3.0 28.3) Strong cybersecurity capability (CyberGym 84.5%) 1M-token context window Efficient token usage vs closed flagships

What are the weaknesses of this model?

Weights not published at launch (staged ~2 weeks later) Trails closed flagships on DeepSWE Thinking cannot be disabled (low/high/max only)

What are the best use cases?

Long-horizon software engineering & repo-level coding Source-code security auditing & vuln discovery Autonomous coding agents Enterprise private-deploy code assistants

Z.aiOpen Source

GLM-5.3

Name: GLM-5.3
Author: Z.ai

Compare this model

Z.ai's (formerly Zhipu AI) GLM-5.3 is a 743B-parameter open-weight MoE flagship released Aug 14, 2026. Rather than scaling pre-training, Z.ai pushed post-training on the unchanged GLM-5.2 base, lifting it to the top open model on agentic coding and cybersecurity: Terminal-Bench 3.0 28.3, DeepSWE v1.1 66.9, CyberGym 84.5%. Weights ship under MIT roughly two weeks after launch.

Share:X Hatena LINE

Parameters

743B (MoE, ~40B active)

Context Window

License

MIT (weights pending)

Release Date

2026-08-14

Benchmark Performance

AA Intelligence Index

—

LMArena Elo

—

HLE

62.5

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

API pricing for this model is not yet available

Strengths

•SOTA open-weight agentic coding (Terminal-Bench 3.0 28.3)
•Strong cybersecurity capability (CyberGym 84.5%)
•1M-token context window
•Efficient token usage vs closed flagships

Weaknesses

•Weights not published at launch (staged ~2 weeks later)
•Trails closed flagships on DeepSWE
•Thinking cannot be disabled (low/high/max only)

Use Cases

•Long-horizon software engineering & repo-level coding
•Source-code security auditing & vuln discovery
•Autonomous coding agents
•Enterprise private-deploy code assistants

Deep Analysis

Context Window

1M tokens

128K max output; text input; built for long-horizon agent work

API Price

$1.40 / $4.40 per 1M

Z.ai list price; Chinese list ~8/28 yuan per 1M; GLM-5.3-Flash ~10x cheaper

SWE-bench Verified

95.4

Independently run; vendor-cited

Terminal-Bench 3.0

28.3

Up from 4.6 on GLM-5.2; vendor-reported

Intelligence Index

60 (Artificial Analysis)

Tied with Kimi K3 as top open-weight in this index

License

GLM-5.3 License

Open weights, but MoA operators above $10B revenue need Z.ai security review

Strengths

・Open weights at 753B params make it the strongest self-hostable frontier-class model alongside Kimi K3.
・Agentic-coding jump (Terminal-Bench 3.0 28.3, DeepSWE 66.9) comes from post-training, not a new pretrain, so the base stays stable.
・$1.40/$4.40 undercuts frontier coding models that score similarly.
・1M context with 128K output suits long-horizon agent and security-research work.

Weaknesses

・Custom GLM-5.3 License (not MIT) imposes a security-review hurdle on large Model-as-a-Service operators.
・Most benchmark scores are vendor-reported except the independent SWE-bench figure.
・Emergent cybersecurity scores (CyberGym 84.5, ExploitBench 54.4) raise dual-use concerns self-hosters must govern.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GLM-5.3	95.4 (SWE-bench)	28.3 (Terminal-Bench 3.0)	88.1	$1.40/$4.40
DeepSeek V4.1 Flash	N/A	N/A	N/A	$0.30/$1.20 (open)
Qwen3.8-Max-0902	N/A	N/A	N/A	$2/$6
Claude Fable 5.1	N/A	N/A	N/A	$10/$50

Overview

GLM-5.3 is Zhipu AI 753B open-weight model released August 14, 2026, built by extended post-training on the GLM-5.2 base rather than a new pretrain. It pairs a 1M-token context with strong agentic-coding scores and a $1.40 per 1M input price, putting frontier-class reasoning within reach of self-hosted stacks.

Benchmarks & Performance

Vendor-reported figures place it at 95.4 on SWE-bench Verified (an independently run number), 88.1 GPQA Diamond, 86.8 MMLU-Pro, 62.5 HLE with tools, and 28.3 on Terminal-Bench 3.0, up from 4.6 on GLM-5.2. The standalone Artificial Analysis Intelligence Index is 60, tied with Kimi K3 as the strongest open-weight models in that index.

Detailed Comparison

Against DeepSeek V4.1 Flash it trades a cheaper open-weight alternative for materially higher coding and reasoning scores. Against Claude Fable 5.1 it gives up some top-end Intelligence Index for roughly 7x lower output cost and full self-hosting. The differentiator is open weights plus frontier quality that closed labs usually reserve for themselves.

Community Feedback

Reaction centered on the license twist: open weights, but a Z.ai security review for Model-as-a-Service operators above $10B revenue, read by some as a hedge against the distillation pressure Anthropic keeps complaining about. The 753B weights landed ungated on Hugging Face and Chinese enterprises adopted it fast for in-house agent stacks.

Use Cases

Best fit is long-horizon agentic coding and security research inside a controlled perimeter: refactoring large repos, exploit reasoning, and tool-driven workflows where a 1M context and 128K output remove the chunking tax. Teams above the $10B revenue threshold should budget for the license review, and treat the self-reported cybersecurity scores as a governance signal rather than a green light.

Latest News

Zhipu released GLM-5.3 on August 14, 2026 via the Z.ai API and Coding Plan, with weights on Hugging Face (zai-org/GLM-5.3) days later. A companion GLM-5.3-Flash (320B/18B multimodal, MIT) followed on August 26 at roughly a tenth of the price.

Sources

Analysis generated: 2026-09-19