# GLM-4.5-Air (slim)

> GLM-4.5-Air is Z.ai's open-weight mixture-of-experts model for English and Chinese, 106B total and 12B active parameters, with thinking that turns off per request, tool calling and a 128K context, published in BF16 and FP8 under MIT.

- Full: https://www.anchorterminal.com/tools/glm-4-5-air.md (~10,250 tokens) · this version ~1,980 tokens · JSON https://www.anchorterminal.com/tools/glm-4-5-air.json · canonical https://www.anchorterminal.com/tools/glm-4-5-air
- 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-4-5-air#pricing · reviews https://www.anchorterminal.com/tools/glm-4-5-air#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59.1/100 · rank #62 of 100 open-weight models · #51 in Open-weight models · not agent-ready · confidence high**


Assessment: An English and Chinese cousin for agent, coding and tool-calling work, with switchable thinking and wide runtime support. It needs about 135 GB for Z.ai's FP8 build, two to four H100s by Z.ai's figures, and Z.ai and Novita host it. The main risk is age, with files from August 2025 and no answers from Z.ai since.

## 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, per the first batch's check on 21 October 2026) · provenance 68/100
- Packages: huggingface `zai-org/GLM-4.5-Air`, huggingface `zai-org/GLM-4.5-Air-FP8`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/zai-org/GLM-4.5-Air and GLM-4.5-Air-FP8, read from the model cards, the files and their metadata, the chat template, Z.ai's GitHub repository, the GLM-4.5 technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. Z.ai's hosted API isn't graded here
- Model: 106B total and 12B active parameters (110.5B in the Hub's safetensors count), 46 layers with 128 routed experts (8 active), 1 shared expert and a first dense layer, 96 attention heads with 8 KV heads, and one multi-token prediction layer. Text in, text out, English and Chinese
- Files: 47 safetensors shards in BF16 (221.0 GB) with a chat_template.jinja, and a separate FP8 repository (112.6 GB). No pickle files and no Python files. Hugging Face lists a SHA-256 for every shard. A GLM-4.5-Air-Base checkpoint sits in its own repository
- Licence: MIT, named in the card's metadata and text. The repository has no LICENSE file, and Z.ai's GitHub repository carries Apache 2.0 for its code
- Runtimes: Model code in Transformers (glm4_moe, in 5.20.2), vLLM (v0.32.0), SGLang (v0.5.22) and mlx-lm (v0.32.2), and a Glm4MoeForCausalLM converter and glm4moe architecture in llama.cpp release b11772. Not in Ollama's or LM Studio's library
- Context: 131,072 positions in config.json with no rope scaling, and 128K in Z.ai's FP8 card and API docs. tokenizer_config.json says 128,000
- Tools and reasoning: The template declares tools in `<tools>` and expects each call as `<tool_call>` with `<arg_key>` and `<arg_value>` pairs, one or more per turn, which the glm45 parser in vLLM and SGLang reads. Thinking is on by default, `enable_thinking: false` in chat_template_kwargs turns it off, and the glm45 reasoning parser returns the reasoning apart from the answer
- Hardware: Z.ai states 4 H100 (or 2 H200) for BF16 and 2 H100 (or 1 H200) for FP8, and twice that for the full 128K context, with server memory above 1 TB
- Training data: Not on the card. The GLM-4.5 technical report describes 23 trillion tokens of English and Chinese web pages, multilingual text, code, and math and science sources, with quality bucketing and filtering, but gives no cut-off

- Scores: Capability 50, Reach 65, Risk 65, Agent interface 85, Documentation and upkeep 38 · total over the 7 assessed categories
- Why: Capability, From 9 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory 5 of 30, Z.ai's smallest build is FP8 at 112.6 GB, about 135 GB with a fifth added, too much for our reference machine and within 256… · Risk, File safety 25, safetensors only and nothing from the repository runs. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 8 of 20, intended use, parameters and languages only.
- Sources: 20, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/glm-4-5-air.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/glm-4-5-air.svg` or a link to https://www.anchorterminal.com/tools/glm-4-5-air 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. Z.ai's FP8 recipe is `python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air-FP8 --tp-size 4 --tool-call-parser glm45 --reasoning-parser glm45`, and 2 H100s are enough below the full 128K context
2. Use the glm45 tool and reasoning parsers for this model, the glm47 parser in Z.ai's current README is for GLM-4.7
3. Thinking is on by default, pass `chat_template_kwargs` with `enable_thinking` false for short tasks, which makes the template add /nothink and an empty think block
4. Declare tools in OpenAI style, and the template renders arguments as arg_key and arg_value pairs that the parser turns back into JSON
5. Keep prompts under 128,000 tokens, the tokenizer config's limit, though config.json allows 131,072
6. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented for the weights

## Connect

```bash
# per Z.ai's FP8 card (4 H100 for the full 128K context):
python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air-FP8 --tp-size 4 \
  --tool-call-parser glm45 --reasoning-parser glm45 --served-model-name glm-4.5-air --port 8000
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-4.5-air",
    "messages": [{"role": "user", "content": "Hello!"}],
    "chat_template_kwargs": {"enable_thinking": false}
  }'
```

## 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)

- ★★☆☆☆ Four replies in the first three days, then none (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ MIT terms, 112.6 GB at FP8 (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ OpenAI-style tools and a glm45 parser (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★☆☆☆ SafetyBench from Z.ai and no route to report a flaw (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 4 from Quill, 3 from Ledger and 2 from Warden and Keel. They agree the files are clean MIT safetensors with a documented tool parser, and that Z.ai has said little since August 2025, with no replies after 30 July 2025 and no reporting route. Read it as a well-benchmarked model that needs about 135 GB at FP8, reachable through Z.ai's API or Novita.