What it is
GLM-5.3 is the flagship model from Z.ai (formerly Zhipu AI), listed in Z.ai's release notes on 18 August 2026, with weights on Hugging Face from 28 August. It is a 753-billion-parameter mixture-of-experts text model (256 experts, 8 used per token) on the same base as GLM-5.2, with all gains from further training. Z.ai ships the weights natively in FP8, alongside a BF16 copy. A multimodal sibling, GLM-5.3-Flash (321B parameters), reads images as well as text.
It has a 1M-token context and up to 128K output tokens. Reasoning is always on, at low, high or max effort, and defaults to max. Z.ai also highlights an unplanned rise in cyber-offence ability: it reports state-of-the-art CyberGym results and more than double GLM-5.2's scores on exploitation benchmarks.
The weights use a modified MIT licence. Only companies running a model-as-a-service business with more than $10 billion in yearly revenue must pass a Z.ai security review before commercial use. On Z.ai's API it costs $1.40 per million input tokens and $4.40 per million output tokens, with cached input at $0.26.
Our review
GLM-5.3 is the strongest open-weight model for coding and agent work that a business can deploy without negotiation. Artificial Analysis gives it 45 on its Intelligence Index, third of 121 models in its open-weights comparison, against 52 for GPT-6.1 Sol and 58 for Claude Opus 5.5. On Z.ai's own table it scores 88.2 on Terminal-Bench 2.1 and 66.9 on DeepSWE v1.1, up from 81.0 and 46.2 for GLM-5.2, and its GDPval-AA v2 rating of 1769, measured by Artificial Analysis, tops every model in that comparison.
Independent human ratings agree it is close but not leading. In the Surge AI blind coding evaluation Mistral published with Large 4, GLM-5.3 scored 3.60 out of 5, level with Kimi K3 (3.59) but behind Mistral Large 4 Preview (3.74) and well behind Claude Opus 5 (4.22). Kimi K3 also edges it on DeepSWE and Toolathlon in Z.ai's own numbers.
The headline price flatters it. At $1.40 and $4.40 per million tokens it is a fraction of closed frontier rates, but Artificial Analysis found it produced 210 million output tokens across its index, against a median of 140 million, and costs $2.01 per index task. That is cheaper than GPT-6 Astra's $3.26 but nearly three times GPT-6.1 Sol's $0.72, because reasoning defaults to max and cannot be switched off.
Security teams should note what Z.ai itself flags. The card describes cyber capability that grew faster than expected, with ExploitGym scores of 105 and 130 against 29 and 39 for GLM-5.2, and the open weights carry no usage restrictions beyond the revenue clause. That makes it useful for authorised security research and a model whose deployment, unlike a closed API, comes with no provider-side safeguards at all.
Where it earns its place
A staff engineer running coding agents in Claude Code or Cline
Cut agent spend without dropping to a weak model
Point the agent at Z.ai's API or the GLM Coding Plan. A month of 100M input tokens (80% cached) and 20M output costs about $137 at list rates, but budget for long reasoning traces at the default max effort.
An ML infrastructure lead at a bank that cannot send code to third-party APIs
Self-host a near-frontier coding model on premises
The FP8 checkpoint is about 753 GB, more than an 8×80 GB node holds, so plan for 141 GB-class GPUs or two nodes with vLLM or SGLang. The licence needs no approval below $10 billion revenue.
A security research team at a pentest firm
Automate vulnerability discovery in authorised engagements
Z.ai reports 84.5% on CyberGym, the top score in its table, and the open weights let the team run it air-gapped. They also own the misuse controls, since nothing upstream filters requests.
Before you commit
- Setup
- Z.ai API is OpenAI-compatible and works in Claude Code, Cline and OpenCode; self-hosting has recipes for vLLM, SGLang, KTransformers and Unsloth, plus Huawei Ascend.
- Integrations
- Function calling, streaming tool calls, structured output, context caching and web search on Z.ai's API; OpenRouter lists it at $0.70/$2.20 per 1M tokens.
- Data and privacy
- Z.ai's API is run from Singapore and its data processing terms say it processes customer data on the customer's behalf; a specific statement on training with API prompts is not published. Self-hosting keeps data in-house.
- Cost at scale
- Z.ai lists $1.40/$4.40 per 1M tokens with cached input at $0.26 and cache storage free for a limited time. Artificial Analysis puts it at $2.01 per index task, against $0.72 for GPT-6.1 Sol. Self-hosting means about 753 GB of FP8 weights plus cache memory.
- Limits
- Text only, reasoning always on (default max), 1M context and 128K output. All benchmark figures except GDPval-AA and the Surge evaluation are Z.ai's own.
What's good
- Modified MIT licence; restriction only above $10B revenue
- 1M-token context, 128K output
- Terminal-Bench 2.1 at 88.2, per Z.ai
- Cheap multimodal Flash sibling at $0.15/$0.50
What's not
- Very verbose: reasoning cannot be switched off
- Main model is text-only
- FP8 weights alone are about 753 GB
- Strong offensive-cyber ability with no built-in safeguards
Best for
Developers and companies who want a capable open coding and agent model with minimal licence restrictions.
Alternatives
GLM-5.3: common questions
- Is GLM-5.3 worth it?
- Open-weight coding within a few points of the closed frontier, under a licence almost nobody needs to read twice. We rate it 4.4 out of 5.
- How much does GLM-5.3 cost?
- Open weights under a modified MIT licence. Z.ai API: $1.40 input / $0.26 cached / $4.40 output per 1M tokens; GLM-5.3-Flash $0.15 / $0.50.
- What are the downsides of GLM-5.3?
- Very verbose: reasoning cannot be switched off; main model is text-only; fP8 weights alone are about 753 GB; strong offensive-cyber ability with no built-in safeguards.
- Who is GLM-5.3 best for?
- Developers and companies who want a capable open coding and agent model with minimal licence restrictions.
- How much does GLM-5.3 cost for a team or in production?
- Z.ai lists $1.40/$4.40 per 1M tokens with cached input at $0.26 and cache storage free for a limited time. Artificial Analysis puts it at $2.01 per index task, against $0.72 for GPT-6.1 Sol. Self-hosting means about 753 GB of FP8 weights plus cache memory.
- Is GLM-5.3 safe for company data?
- Z.ai's API is run from Singapore and its data processing terms say it processes customer data on the customer's behalf; a specific statement on training with API prompts is not published. Self-hosting keeps data in-house.
- What are the best GLM-5.3 alternatives?
- The closest alternatives we have reviewed are Kimi K3, DeepSeek-V4-Pro, Qwen3.8-Max and MiniMax-M3.
Answers checked Oct 23, 2026.
Make GLM-5.3? Show this rating on your site
