AI Models Navi
모델 목록으로
Zhipu AI오픈소스

GLM-5

GLM-5 is a large-scale foundation model developed by Zhipu AI. It has an extremely large scale with approximately 7.4 trillion parameters and is a chat-specific model that supports a long context window of 200K.

파라미터

7440.0B

컨텍스트

200K

라이선스

MIT

출시일

2026-02-11

API 가격

입력 가격 (1M 토큰당)

$1

출력 가격 (1M 토큰당)

$

과금 모드: standard

강점

  • ・Over 7.4-trillion massive parameters
  • ・Supports long 200K context
  • ・High openness via MIT license

약점

  • ・Extremely large 1.51TB file size
  • ・Requires very high compute resources to run
  • ・Inference cost scales with model size

활용 사례

  • ・Advanced conversational AI chatbots
  • ・Ultra-long document analysis and summarization
  • ・R&D for large foundational models

심층 분석

Parameters

744B total / 40B active MoE

Arena Elo

1452-1456

#1 open-weight

SWE-Bench Verified

77.8%

AIME 2025

92.7%

Pricing

$1.00/$3.20 per 1M tokens

License

MIT

Training Hardware

Huawei Ascend 910B

Release Date

February 11, 2026

강점

  • ・#1 open-weight on LMArena (1452 Elo)
  • ・SWE-Bench 77.8% beats Gemini 3 Pro and GPT-5.2
  • ・MIT license
  • ・Trained on Huawei Ascend (zero Nvidia)
  • ・Strong agentic performance (TAU-Bench 98%)

약점

  • ・Text-only, no multimodal
  • ・744B requires 8x H200 for self-hosting
  • ・GPQA trails frontier by 5-10 points

경쟁사 비교

ModelArenaSWEGPQAPrice
Claude Opus 4.6154857.3 Pro91.3$5/$25
GPT-5.275.4$2.50/$10
Gemini 3 Pro~144976.2$1.25/$5.00

GLM-5 is Zhipu AI's flagship 744B MoE model. #1 open-weight on LMArena (1452 Elo), 77.8% SWE-Bench Verified surpassing Gemini 3 Pro and GPT-5.2. Trained entirely on Huawei Ascend chips, MIT licensed.

출처

분석 생성일: 2026-05-24