DEV.co
Open-Source LLM · zai-org

GLM-5

GLM-5 is a 744-billion-parameter open-source LLM developed by zai-org, released in April 2026. It uses sparse mixture-of-experts (MoE) architecture with 40B active parameters and integrates DeepSeek Sparse Attention to reduce deployment cost. The model is designed for complex reasoning, coding, and agentic tasks. It supports both English and Chinese, is MIT-licensed, ungated, and can be self-hosted using multiple open-source frameworks.

Source: HuggingFace — huggingface.co/zai-org/GLM-5
753.9B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
64.6k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerzai-org
Parameters753.9B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads64.6k
Likes2.1k
Last updated2026-04-05
Sourcezai-org/GLM-5

What GLM-5 is

GLM-5 is a 744B-parameter MoE transformer trained on 28.5T tokens with post-training via asynchronous RL (slime infrastructure). It uses DeepSeek Sparse Attention (DSA) to enable long-context capacity with reduced VRAM footprint. The model card lists support for vLLM, SGLang, KTransformers, Hugging Face Transformers, and xLLM. Context length is not specified. Distributed inference with tensor-parallel-size 8 is recommended.

Quickstart

Run GLM-5 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="zai-org/GLM-5")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Software Engineering & Code Tasks

SWE-bench Verified 77.8%, SWE-bench Multilingual 73.3%. Suitable for code-generation pipelines, bug fixing, and multi-file refactoring tasks requiring long context.

Agentic Systems & Tool Use

Strong performance on HLE (30.5), Terminal-Bench 2.0, and BrowseComp (75.9 with context management). Designed for complex systems engineering and long-horizon agentic workflows.

Self-Hosted AI Applications

MIT license, ungated, supports vLLM/SGLang/KTransformers. Ideal for organizations needing on-premises or cloud VPC deployment without licensing friction.

Running & fine-tuning it

ESTIMATE: 744B parameters at FP16 ≈ 1.5 TB raw; with MoE sparsity and DSA, active memory likely 300–500 GB. Recommended deployment: 8× A100/H100 GPUs (80 GB each) with tensor parallelism. vLLM deployment recipe specifies tensor-parallel-size 8, --gpu-memory-utilization 0.85. Quantization (GPTQ/AWQ) may reduce to 100–200 GB. Verify with your inference framework and batch size.

Not stated in model card. No mention of LoRA, QLoRA, or fine-tuning infrastructure. GLM series historically supports instruction fine-tuning. Recommend: (1) check GitHub repo for recipes, (2) expect full fine-tuning to require multi-GPU setup due to scale, (3) assess LoRA feasibility empirically or contact zai-org.

When to avoid it — and what to weigh

  • Tight Latency SLAs on Limited Hardware — Model requires 8-GPU tensor-parallel setup. Even with MoE sparsity, inference cost is substantial. Not suitable for single-GPU or edge deployments without quantization.
  • Production Without Benchmarking Against Your Workload — Published benchmarks are curated; real-world performance on proprietary datasets or specialized domains may differ. Requires custom evaluation before committing.
  • Strict IP/Output Control Requirements — MIT license allows redistribution of model outputs. If you need contractual guarantees on derivative works or cannot tolerate community fine-tuned versions, consult legal.
  • Lack of Commercial Support Need — zai-org offers API services and Z.ai platform. No information on commercial SLAs, support response times, or enterprise contracts in the model card.

License & commercial use

MIT License. Permissive open-source license allowing commercial use, modification, and redistribution with minimal restrictions (attribution required).

MIT is a permissive OSI-approved license. Commercial use of the model weights and outputs is allowed. However: (1) No explicit warranty or indemnity; (2) zai-org offers commercial API services separately (Z.ai), which may have different terms; (3) If you integrate GLM-5 into a product, ensure compliance with MIT attribution and any downstream license obligations. Consult legal for high-stakes deployments.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM risks apply: (1) Model outputs may contain harmful content; use content filtering in production, (2) Ungated and open-source; source-code and weights publicly available—review for any embedded artifacts before deploying, (3) Inference on user data: ensure appropriate isolation, rate-limiting, and audit logging, (4) Tool-calling/agentic capability increases attack surface—validate tool specifications and sandbox execution, (5) No security disclosure policy mentioned; file responsibly if vulnerabilities found. No formal adversarial robustness testing cited.

Alternatives to consider

DeepSeek-V3.2

Also uses MoE with sparse attention (DeepSeek Sparse Attention itself). 73.1% SWE-bench vs GLM-5's 77.8%; comparable reasoning. Unclear license and commercial terms; requires review.

Llama 3.3 (Meta)

Open-source, widely deployed, strong on coding (SWE-bench). Llama 3.3 is Llama 2 license (requires review for commercial use). Smaller/less capable on long-context agentic tasks than GLM-5.

Grok-3 or Claude Opus 4.5 (proprietary)

Comparable or better reasoning/agentic performance (Claude achieves 80.9% SWE-bench). No deployment overhead, but closed-source, vendor lock-in, higher per-token costs, no fine-tuning control.

Software development agency

Ship GLM-5 with senior software developers

Start with vLLM or SGLang on 8× A100/H100 GPUs. Benchmark against your workload before production. For managed inference, use Z.ai API Platform. For custom fine-tuning or on-premises deployment, contact zai-org or consult the GitHub repository.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

GLM-5 FAQ

Can I use GLM-5 commercially without restrictions?
MIT license permits commercial use, modification, and redistribution. No explicit license restriction. However, zai-org separately offers commercial API services (Z.ai); if you self-host, ensure you have infrastructure/support in place. For high-stakes products, have legal review the MIT terms and any downstream liability allocation.
What is the minimum GPU setup to serve GLM-5?
Not explicitly stated, but vLLM deployment recipes recommend tensor-parallel-size 8 (e.g., 8× A100/H100 GPUs). With aggressive quantization and single-model batching, smaller setups may work; test empirically. Running on fewer GPUs will incur sequence-parallel or pipeline-parallel complexity not documented here.
Does GLM-5 support fine-tuning, and if so, what framework?
Not documented in the model card. GitHub repo likely has recipes. Historically, GLM models support instruction fine-tuning; expect multi-GPU setups. LoRA/QLoRA feasibility unknown—test or contact zai-org for guidance.
How long is the context window?
Unknown. Benchmark footnotes reference 200K, 250K, and 128K context for different evals, suggesting variable capabilities. Check paper (arxiv:2602.15763) or GitHub for definitive context length.

Software development & web development with DEV.co

Need help beyond evaluating GLM-5? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy GLM-5?

Start with vLLM or SGLang on 8× A100/H100 GPUs. Benchmark against your workload before production. For managed inference, use Z.ai API Platform. For custom fine-tuning or on-premises deployment, contact zai-org or consult the GitHub repository.