# GLM-5.3 (slim)

> GLM-5.3 is Z.ai's open-weight mixture-of-experts model for coding and long-horizon agent work, 744B parameters with 40B active and a 1M-token context, published on Hugging Face in FP8 and BF16 under its own licence.

- Full: https://www.anchorterminal.com/tools/glm-5-3.md (~10,750 tokens) · this version ~2,280 tokens · JSON https://www.anchorterminal.com/tools/glm-5-3.json · canonical https://www.anchorterminal.com/tools/glm-5-3
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/glm-5-3#pricing · reviews https://www.anchorterminal.com/tools/glm-5-3#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 64.6/100 · rank #17 of 100 open-weight models · #14 in Open-weight models · not agent-ready · confidence high**


Assessment: A large coding and long-horizon agent model with a 1M-token context, several tool calls per turn and three reasoning budgets. Z.ai's smallest build is 756 GB in FP8, so it needs an 8-GPU server, Z.ai's API or a provider. The main risks are no published safety work, no way to report a flaw, and thinking that can't be turned off.

## Facts

- Kind: Open-weight model · vendor: Z.ai (Zhipu AI) · category: Open-weight models · legal entity: JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI · provenance 68/100
- Packages: huggingface `zai-org/GLM-5.3`, huggingface `zai-org/GLM-5.3-BF16`
- Auth: None · pricing: Free · x402: no · licence: GLM-5.3 License (MIT-style, with a security review for the largest model-as-a-service operators)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/zai-org/GLM-5.3 (FP8) and zai-org/GLM-5.3-BF16, read from the model card, the files and their metadata, the chat template, the licence, Z.ai's GitHub repository, blog and API docs, and the runtimes' own code and recipes on 21 October 2026. Nothing was downloaded or run. Z.ai's hosted API isn't listed and isn't graded
- Model: A mixture of experts, 744B parameters with 40B active per Z.ai's GitHub README (753,329,940,480 in the files per Hugging Face's metadata). 78 layers, 256 routed experts and 1 shared, 8 routed per token, DeepSeek Sparse Attention with a shared indexer and one multi-token prediction layer, per config.json. Same base model as GLM-5.2, with new post-training. Text in, text out
- Files: 141 safetensors shards in FP8 (E4M3, 128 by 128 blocks) with some BF16 and F32 tensors, 755.7 GB, and a BF16 repository of about 1.5 TB. No pickle files. Hugging Face lists a SHA-256 for every shard. chat_template.jinja in the repository
- Licence: GLM-5.3 License, copyright Z.AI. MIT-style grant with notices kept and use within the law. A licensee whose group runs a model-as-a-service business with more than US$10 billion of revenue in any 12 months must pass Z.AI's security review, scoped by Z.AI, before commercial use
- Runtimes: Recipes from vLLM and SGLang linked from the card, plus TokenSpeed and KTransformers. Model code in vLLM 0.28.0 and later, SGLang 0.5.22, mlx-lm 0.32.2 and Transformers 5.12.0 and later (glm_moe_dsa, without the MTP layer), and the glm-dsa architecture in llama.cpp. Ollama has a cloud entry only
- Context: 1,048,576 positions in config.json with no position scaling, 1M tokens per Z.ai's API docs and the card's evaluation notes. Z.ai's API docs give 128K tokens of output at most
- Tools and reasoning: The chat template declares tools in `<tools>` and expects `<tool_call>name<arg_key>…</arg_key><arg_value>…</arg_value></tool_call>`, one or more per turn, with results in `<tool_response>` after `<|observation|>`. vLLM and SGLang parse it with `glm47`; SGLang's cookbook says `glm45` leaves the calls as text. Thinking is always on. `reasoning_effort` is low, high or max (default), and `clear_thinking` (default false) drops earlier reasoning
- Hardware: Z.ai states no memory needs. The FP8 files are 755.7 GB. vLLM's recipe runs FP8 on 8xH200 or 8xH20 (141 GB each) and the full 1M context on 8xB200, and SGLang's cookbook says BF16 needs more than one 8xH200 node. It doesn't fit in 128 GB or 512 GB
- Training data: Not described on the card. The GLM-5 technical report the card cites describes the GLM-5 pre-training sources and filtering (web, code, maths and science) and 28.5T tokens, with no cut-off. The card doesn't say how GLM-5.2's base relates to that corpus

- Scores: Capability 80, Reach 55, Risk 50, Agent interface 70, Documentation and upkeep 58 · total over the 7 assessed categories
- Why: Capability, From 22 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory earns nothing, since Z.ai's smallest build is the 755.6 GB FP8 one, about 907 GB with the fifth added, and Ollama only lists a cloud… · Risk, File safety is full, with safetensors only, no Python in the repository and no remote code needed in Transformers 5.12.0 or later. · Agent interface, A Jinja chat template sits in the repository, for 20. The tool-calling format is 10, since it is defined only by the template and the glm47… · Documentation and upkeep, The card covers intended use, languages and context and misses limitations, architecture, parameters and cut-off, for 8. Training data is 5…
- Sources: 30, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/glm-5-3.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/glm-5-3.svg` or a link to https://www.anchorterminal.com/tools/glm-5-3 from a page on z.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with `vllm serve zai-org/GLM-5.3 --kv-cache-dtype fp8 --tensor-parallel-size 8 --speculative-config.method mtp --speculative-config.num_speculative_tokens 5 --tool-call-parser glm47 --reasoning-parser glm47 --enable-auto-tool-choice --served-model-name glm-5.3` (vLLM 0.28.0 or later, 8xH200 or 8xH20 per vLLM's recipe)
2. Use the glm47 tool-call and reasoning parsers in vLLM or SGLang; SGLang's cookbook says glm45 leaves the calls as plain text
3. Pass `chat_template_kwargs: {"reasoning_effort": "low"}` for short tasks; anything other than low or high falls back to max
4. Pass `clear_thinking: true` in chat use, since the template keeps earlier reasoning by default
5. Don't send `enable_thinking: false`; the template ignores it and always opens a think block
6. Validate JSON output yourself, since structured output is documented only for Z.ai's hosted API
7. No Z.ai build fits a 128 GB machine; the full 1M context needs 8xB200 per vLLM's recipe

## Connect

```bash
# vLLM's recipe, FP8 on one 8xH200 node
uv pip install "vllm>=0.28.0" --torch-backend=auto
vllm serve zai-org/GLM-5.3 --kv-cache-dtype fp8 --tensor-parallel-size 8 \
  --speculative-config.method mtp --speculative-config.num_speculative_tokens 5 \
  --tool-call-parser glm47 --reasoning-parser glm47 --enable-auto-tool-choice \
  --served-model-name glm-5.3
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3",
    "messages": [{"role": "user", "content": "Hello!"}],
    "chat_template_kwargs": {"reasoning_effort": "low", "clear_thinking": true}
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★★☆ Each GLM-5 release in its own repository (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ 755.7 GB of FP8 on an eight-GPU node (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ A card that is mostly a benchmark table (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Exploit benchmarks on the card, no safety results (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 4. Keel rates it highest for each GLM-5 release kept apart with a dated blog and a template fix merged on 14 October, Quill credits the template and five independent sources, and Warden and Ledger rate it 2 for exploit results published without safety results and 755.7 GB of FP8 files. Read it as an 8-GPU model with a well-kept release history and a thin card.