Z

GLM-5.1

China's Open-Source Assault on Claude Opus

Z.ai (Zhipu AI)  •  March 2026  •  MIT Licensed

3x Credits vs Opus Open Source 744B MoE

Who is Z.ai?

  • Formerly Zhipu AI — rebranded to Z.ai
  • Founded at Tsinghua University, Beijing
  • Hong Kong IPO — Jan 8, 2026, raised ~$558M
  • First publicly traded foundation model company
  • On US Entity List since 2025
  • Trains entirely on Huawei Ascend chips — zero NVIDIA
  • GLM family since 2022 — 4 years of open-weight LLMs
GLM-130B 2022 GLM-4 2023-24 GLM-4.5/4.7 2025 GLM-5 Feb 2026 GLM-5.1 Mar 2026 HK IPO Jan 2026 $558M raised US Entity List Huawei Ascend only

GLM-5: The Foundation

744B
Total Parameters
40B
Active per Token (MoE)
200K
Context Window
28.5T
Pretraining Tokens
  • Released Feb 11-12, 2026 by Z.ai
  • MoE architecture: 256 experts, top-8 activated
  • DeepSeek Sparse Attention (DSA) — linear complexity
  • "Slime" async RL framework for post-training
  • MIT license — weights on HuggingFace
  • API: ~$1/M input tokens
  • Trained on Huawei Ascend — no NVIDIA needed
Input Token Router (top-8 of 256) Expert Expert Active Expert Expert 256 experts total Only 8 active = 40B params Sparsity ~5.9%

GLM-5 Benchmark Performance

77.8%
SWE-bench Verified
$4,432
Vending Bench 2
56-61%
Terminal-Bench
GLM-5
77.8%
Claude Opus 4.6
~72%
GPT-4o
~65%
DeepSeek V3.2
~70%

SWE-bench Verified measures real-world GitHub issue resolution • Higher = better

Why Developers Loved GLM-5

  • MIT License — self-host, modify, commercialize freely
  • Competitive coding at fraction of closed model cost
  • Strong tool-use — built for agent harnesses
  • CodeGeeX4 coding variant — 2M+ VS Code installs
  • HN community: "Finally an open model that can do real agent work"
  • Windsurf & opencode users: GLM as cheap Opus replacement for high-volume iteration
  • $1/M tokens vs Opus $15/M — 15x cheaper
GitHub Stars Growth Post-Launch 🔥 CodeGeeX4 2M+ VS Code installs

GLM-5-Turbo: The Pivot That Scared Everyone

Mar 15-16, 2026
Z.ai releases GLM-5-Turbo — closed-source, API-only
"No commitment to open-sourcing Turbo"
  • Lower tool-call error rates
  • Faster throughput for agent tasks
  • Built for Z.ai's own agent ecosystem
  • Community reaction: panic
  • "Is Z.ai abandoning open source?"
  • Context: Alibaba Qwen 3 exec departures in 2026
    Questions about open-model profitability
GLM-5 MIT Open GLM-5-Turbo Closed API Community feared this shift → was permanent

GLM-5.1: The Response

@ZixuanLi_ (Li Zixuan, Global Head Z.ai)
"Don't panic. GLM-5.1 will be open source."
March 20, 2026
✓
Open Source Confirmed
Mar 27
Released 2026
3x
Credits vs Claude Opus
Direct response to community backlash over GLM-5-Turbo. Z.ai publicly commits: next flagship stays MIT licensed.

The Price War: GLM-5.1 vs Claude Opus

$1
GLM-5.1 per 1M tokens
$15
Claude Opus per 1M tokens
15x
Cheaper than Opus
3x
More credits per dollar
GLM-5.1
$1/M
Claude Opus 4.6
$15/M
GPT-4o
$10/M
Gemini Pro
$6/M
Z.ai marketing: "3x the work per dollar vs Claude Opus" — same benchmark performance, fraction of the cost. MIT licensed: self-host for $0.

GLM-5.1: What's New vs GLM-5

  • Improved tool-call accuracy — lessons from Turbo
  • Enhanced agentic reasoning for long-horizon tasks
  • MIT license maintained — open weights
  • Available: HuggingFace zai-org/GLM-5.1
  • Also on ModelScope: ZhipuAI/GLM-5.1
  • Recommended use: coding agents, long-horizon tasks, cost-sensitive deployments
  • FP8 quantized version available for efficient inference
GLM-5 Feb 2026 • MIT GLM-5-Turbo Mar 15 • Closed GLM-5.1 Mar 27 • MIT • Improved Better tools + agents + stays open source ✓

Chinese AI Landscape 2026

Z.ai (Zhipu)
GLM-5.1 • MIT
744B MoE • 200K ctx • Strong: Coding/Agents • $1/M
DeepSeek
V3.2 • MIT
~700B MoE • 128K ctx • Strong: Reasoning/Math • $0.50/M
Alibaba
Qwen 3.5 • Apache 2.0
Various sizes • 1M ctx • Strong: Multilingual
MiniMax
M2.7 • Proprietary
Unknown params • 1M ctx • Strong: Long context
Moonshot (Kimi)
Kimi 2.5 • Proprietary
Unknown • 2M ctx • Strong: Ultra-long context
ByteDance
Doubao-Seed-2.0 • Proprietary
Unknown • 256K ctx • Strong: Coding

Green = Open Source (MIT/Apache)   Orange = Closed/Proprietary

The Open Source Arms Race

  • Jan 2025 — DeepSeek R1 moment: Chinese open-weight shocks Western labs
  • Feb 2026 — GLM-5 pushes SWE-bench to 77.8% open-weight
  • Chinese labs now leading open-weight frontier
  • Western labs (OpenAI, Anthropic) remain mostly closed
  • MIT licensing = standard for Chinese frontier models
  • Self-hosting: no API lock-in, no censorship, no rate limits
  • Z.ai public commitment: stay open despite commercial pressure
China 🇨🇳 GLM DeepSeek Qwen USA 🇺🇸 Anthropic OpenAI Open Source Closed

Community Reaction

Hacker News
“Finally an open model that can actually do real agent work. GLM-5 changes the calculus for self-hosted deployments.”
opencode community (HN)
“$10/month gets you a decent allowance with latest GLM, MiniMax, and Kimi. Forget $20/month Windsurf for high-volume work.”
2M+
CodeGeeX4 VS Code Installs
#1
GitHub Trending at Launch
15x
Cheaper than Opus for Devs

Should You Use GLM-5.1?

USE IT IF...

  • Coding/agentic tasks at scale
  • Cost matters — $1 vs $15/M tokens
  • Want self-hosted open weights
  • Building agent harnesses
  • MIT freedom to modify and deploy
CONSIDER ALTERNATIVES IF...

  • Need 1M+ token context (Kimi 2.5 or Opus)
  • Multimodal is critical (GPT-4o / Gemini)
  • Inference speed is top priority
  • Need proven enterprise SLA

The Bottom Line

744B
Biggest Open-Weight 2026
3x
Credits vs Claude Opus
MIT
Self-Host for Free
77.8%
SWE-bench Verified
Z.ai built the biggest open-weight model of 2026 and is directly challenging Anthropic on price.
GLM-5.1: same benchmark tier as Claude Opus, 15x cheaper, MIT licensed.
The East is no longer catching up — it is leading in open-weight AI.

HuggingFace: zai-org/GLM-5.1  |  ModelScope: ZhipuAI/GLM-5.1  |  z.ai