AI Agent Hub
Back to models
Z.ai: GLM 5.3 Prime logo

Z.ai: GLM 5.3 Prime

Closed Source z-ai Released 2026-09-23
45.0 / 100 744.0B params 1M context Proprietary

About this model

GLM-5.3-Prime is Z.ai’s high-throughput serving tier built on the same GLM-5.3 foundation as the open-weight release. It inherits the full reasoning, coding, and agentic capabilities of GLM-5.3 while targeting roughly 1.5–2× higher output throughput through inference acceleration, so latency-sensitive agent loops and streaming code generation see faster token delivery without changing the underlying model behavior. The API exposes a one-million-token context window, up to 131K completion tokens, always-on chain-of-thought reasoning with configurable effort levels (low, high, and max), plus tool calling and structured JSON outputs.

On public leaderboards, Prime is evaluated as equivalent to GLM-5.3 because it uses the same weights; official Z.ai and Hugging Face reporting therefore emphasize long-horizon software engineering, terminal agents, and cybersecurity. Representative scores include strong Terminal-Bench 2.1 performance, competitive DeepSWE and SWE-Bench Verified results, and leading open-weight marks on CyberGym and related exploitation benchmarks from the GLM-5.3 post-training push. The model is positioned for multi-turn agent orchestration, repository-scale coding, and security research workflows rather than lightweight chat-only use.

GLM-5.3-Prime is proprietary and available through hosted APIs (for example OpenRouter and Z.ai partners) rather than as downloadable weights. It was released on September 23, 2026, alongside pricing tuned for production agent workloads. Developers migrating from GLM-5.2 or earlier GLM APIs should expect mandatory reasoning, updated chat-template defaults, and the same 744B total-parameter MoE architecture (40B activated per token) as GLM-5.3.

Benchmark Scores

HLE
62.5
IFEval
72.0
DeepSWE
66.9
CyberGym
84.5
MMLU-Pro
86.8
AIME-2026
92.7
GDPval-AA
1769.0
LiveBench
76.14
BrowseComp
75.9
SimpleBench
66.2
GPQA-Diamond
88.1
Context-Arena
88.53
NL2Repo-Bench
58.0
SWE-Bench-Pro
95.6
AutomationBench
48.2
IMO-AnswerBench
82.5
Agents-Last-Exam
28.5
Creative-Writing
2075.0
SWE-Bench-Verified
77.8
Terminal-Bench-2.0
61.1
Terminal-Bench-2.1
88.2
Terminal-Bench-3.0
28.3
Toolathlon-Verified
73.0

Technical Specs

  • Parameters: 744.0B
  • Architecture: MoE Transformer
  • Context Window: 1,000,000 tokens
  • Input Modalities: text

Hardware Requirements

  • Compute: API only

Pricing

Input Output Currency
2.80 / 1M tokens 8.80 / 1M tokens USD