Overview
- Tool name: GLM-5
- Developer: Zhipu AI (Z.ai)
- Official website: https://z.ai/blog/glm-5
- Category: AI Models
- License: MIT (open weights)
GLM-5 is Zhipu AI's open-source flagship: a 744B-parameter MoE model (40B active per token), pretrained on 28.5T tokens, integrating DeepSeek Sparse Attention (DSA) to cut long-context deployment cost. It targets complex systems engineering and long-horizon agentic work - the official framing is "from vibe coding to agentic engineering".
Benchmarks (official): SWE-bench Verified 77.8, Humanity's Last Exam with tools 50.4, GPQA Diamond 86.0, AIME 2026 I 92.7, BrowseComp 62.0. It sits in the top open-source tier across reasoning, coding and agentic tasks, within a few points of closed flagships like Claude Opus 4.5 and GPT-5.2. GLM-5 also carries a historic footnote: the first frontier-class model trained entirely on Chinese hardware (100,000 Huawei Ascend chips).
The ecosystem: open weights on Hugging Face / ModelScope (MIT), APIs on api.z.ai and BigModel.cn, the Z.ai web app (Chat and Agent modes), and a GLM Coding Plan subscription compatible with coding agents like Claude Code and OpenClaw. Successor versions GLM-5.2 / GLM-5.3 are live on the Z.ai platform and GLM models are also served as third-party models on Mistral's platform.
API pricing (reference, per million tokens)
- GLM-5: roughly $1.00 input / $3.20 output, cached input around $0.20 (confirm live rates at api.z.ai)
- 200K context; MIT weights are self-hostable, with quantized builds fitting a single 80GB GPU
Who it is for
- Coding agents: SWE-bench 77.8 leads the open class; Claude Code / OpenClaw compatible as a cheaper alternative
- Long-horizon agents: #1 among open models on Vending Bench 2 (a simulated year of operations), showing strong long-term planning
- Document output: generates finished .docx / .pdf / .xlsx deliverables directly
- Sovereign deployment: MIT license with no additional clauses - the simplest commercial terms in the class
Limitations
- Verbose output (roughly 7x median token consumption in evaluations); budget on actual output volume
- Reasoning-mode speed is moderate (about 48-72 tok/s); watch latency in interactive use
- English-community tooling is newer and less mature than the Llama/Qwen ecosystems
Frequently asked questions
Is GLM-5 open source?
Yes - MIT license, weights on Hugging Face and ModelScope, free for commercial use, fine-tuning and hosting, with no Llama-style additional clauses.
What does GLM-5 have to do with Claude?
Nothing, but GLM-5 is interface-compatible with coding agents like Claude Code and is often used as a cost-effective alternative; Z.ai offers a GLM Coding Plan for exactly this.
What does it cost?
Reference API pricing is $1.00 input / $3.20 output per million tokens (check live rates at api.z.ai); self-hosting costs compute only.
What is the latest version?
After GLM-5, Zhipu has shipped GLM-5.2 and GLM-5.3; GLM-5 (released early 2026) remains the main open-weights release.
*This page is not affiliated with Zhipu AI. Benchmarks are vendor-reported; pricing per api.z.ai. Verified as of October 2026.*
