Overview

GLM-5.3 is a large open-weight language model from Zhipu AI, which sells its models internationally under the Z.ai name. It uses the same base model as GLM-5.2, so all of its gains come from post-training. Z.ai says it is much better at complex coding and long-horizon tasks, and that its cybersecurity capability “developed faster than we expected” as post-training scaled. It reached Z.ai’s GLM Coding Plan in mid-August 2026, Z.ai’s release notes list it on August 18, and the weights followed on Hugging Face later that month after a safety review.

How it works

The Hugging Face configuration describes a mixture-of-experts model (glm_moe_dsa) with about 753 billion parameters in total. Each layer holds 256 routed experts plus one shared expert, and 8 routed experts handle each token. The model has 78 layers and a sparse-attention indexer, and its configured maximum position is 1,048,576 tokens. The model card does not give an active-parameter count. The technical background is the GLM-5 paper (“From Vibe Coding to Agentic Engineering”).

Z.ai’s documentation says the model takes text-only input, with a 1M-token context window and up to 128K output tokens. Reasoning is always on. A reasoning_effort setting chooses low, high or max (the default), and requests that try to disable thinking fail.

Licence, pricing and availability

The weights use the custom GLM-5.3 License. It is permissive in the style of MIT: you may use, modify, sell and fine-tune the weights. The exception is any company that runs a “Model as a Service” business and has more than $10 billion in revenue over 12 months. Such a company must pass Z.ai’s security review before any commercial use. Earlier GLM-5 weights used the plain MIT license.

On the Z.ai API, GLM-5.3 costs $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens. The API accepts the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats. The smaller GLM-5.3-Flash costs $0.15 in and $0.50 out. The model card lists SGLang, vLLM, Transformers, KTransformers, Unsloth and Ascend NPU deployments, plus community quantizations for llama.cpp, Ollama and LM Studio.

Strengths and limits

  • On Z.ai’s own table it scores 88.2 on Terminal Bench 2.1 and 28.3 on Terminal Bench 3.0, up from 4.6 for GLM-5.2. It scores 84.5 on CyberGym, the top result in that table.
  • Rivals beat it on several rows. GPT-5.6 Sol leads Terminal Bench 3.0 with 34.6, and Fable 5 leads ExploitBench.
  • All of these scores are Z.ai’s own. Running the full model locally needs a multi-GPU server.

Alternatives

Other open-weight options include GLM-5, Kimi K3, DeepSeek V4 and Mistral Large 4. Cheaper agent options include MiniMax M3.1 Flash and DeepSeek V4.1 Flash.

Our coverage

Mistral compared Mistral Large 4 with GLM-5.3 on cyber work. A study of coding-agent harnesses included GLM-5.3-Flash among its test models.

Specifications are taken from the model card on Hugging Face and Z.ai’s documentation. Benchmark results depend on evaluation settings; see how AI benchmarks work.