# GLM-5.3-Flash (slim)

> GLM-5.3-Flash is Z.ai's open-weight multimodal mixture-of-experts model, 320B parameters with 18B active and a 1,048,576-token context, published on Hugging Face in FP8 and BF16 under the MIT licence.

- Full: https://www.anchorterminal.com/tools/glm-5-3-flash.md (~10,400 tokens) · this version ~2,080 tokens · JSON https://www.anchorterminal.com/tools/glm-5-3-flash.json · canonical https://www.anchorterminal.com/tools/glm-5-3-flash
- 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-flash#pricing · reviews https://www.anchorterminal.com/tools/glm-5-3-flash#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 65.1/100 · rank #16 of 100 open-weight models · #13 in Open-weight models · not agent-ready · confidence high**


Assessment: An MIT-licensed multimodal agent and coding model with a 1M-token context and several tool calls per turn, served by vLLM, SGLang, llama.cpp and Transformers. Z.ai's smallest build is 328 GB in FP8, so it needs a multi-GPU server. The main risk is that nothing on safety or prompt injection is published and there is no way to report a flaw.

## Facts

- Kind: Open-weight model · vendor: Z.ai · category: Open-weight models · legal entity: Z.AI Co., Ltd (named in the licence's copyright line) · provenance 53/100
- Packages: huggingface `zai-org/GLM-5.3-Flash`, huggingface `zai-org/GLM-5.3-Flash-BF16`
- 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-5.3-Flash (FP8) and zai-org/GLM-5.3-Flash-BF16, read from the model card, the files and their metadata, the chat template, the licence, Z.ai's GLM-5 GitHub repository and developer guide, and the runtimes' own code and recipes on 21 October 2026. Nothing was downloaded or run. Z.ai's hosted glm-5.3-flash API isn't graded here
- Model: A mixture of experts with 320B parameters and 18B active per token, 288 routed experts and 1 shared, 8 routed per token, 45 layers mixing linear attention (34 layers) and sparse attention (11), Manifold-Constrained Hyper-Connections, one multi-token prediction layer and a 24-layer vision encoder. Text, images and video in, text out. 321,323,031,390 parameters in the files per Hugging Face's metadata
- Files: Safetensors in both repositories, 62 shards and 328.4 GB in FP8 (block-wise E4M3, with BF16 for the parts left unquantised) and 120 shards and 642.7 GB in BF16, each with a chat_template.jinja. No pickle files and no Python. Hugging Face lists a SHA-256 for every shard
- Licence: MIT for the weights, copyright Z.AI Co., Ltd. No acceptable use policy in the repository. Z.ai's GLM-5 GitHub repository is Apache 2.0
- Runtimes: Model code in Transformers from 5.16.1 (glm5_next), vLLM 0.32.0, SGLang 0.5.22, llama.cpp (glm5-next, in build b11768) and mlx-vlm 0.7.9. Ollama has a cloud entry only, and mlx-lm has no model file for it. The card links recipes from SGLang, vLLM, TokenSpeed, KTransformers and Unsloth
- Context: 1,048,576 tokens in config.json and the tokenizer config, with no rotary scaling (the attention uses no positional rotation). Z.ai's developer guide gives 1M tokens and 128K output for the API. The card itself doesn't state a context length
- Tools and reasoning: The chat template declares tools in `<tools>` and expects one or more calls as `<tool_call>name<arg_key>…</arg_key><arg_value>…</arg_value></tool_call>`, which vLLM's and SGLang's `glm47` parser reads. Thinking is always on, the generation prompt opens `<think>` every time, and `reasoning_effort` is low, high or max (the default). `clear_thinking` defaults to false, which keeps earlier reasoning in the prompt
- Hardware: The card states no memory needs. The FP8 files are 328.4 GB, vLLM's recipe gives a 386 GB minimum for FP8, and no Z.ai build fits in 128 GB
- Training data: The card and developer guide describe a newly trained base model on a 30T-token multimodal pre-training corpus, with no sources, mix or cut-off

- Scores: Capability 75, Reach 60, Risk 55, Agent interface 70, Documentation and upkeep 58 · total over the 7 assessed categories
- Why: Capability, From 18 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 328.4 GB FP8 one, about 394 GB with the fifth added, and neither Ollama nor LM Stud… · Risk, File safety is full, with safetensors only, no Python and no remote code needed in Transformers 5.16.1 or later. · Agent interface, A Jinja chat template sits in both repositories, for 20. The tool-calling format is 10, because it is defined only by the template and the g… · Documentation and upkeep, The card covers architecture, parameters and languages and misses four of the seven items, for 8. Training data is 5, a 30T-token corpus des…
- Sources: 24, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/glm-5-3-flash.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/glm-5-3-flash.svg` or a link to https://www.anchorterminal.com/tools/glm-5-3-flash 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-Flash --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47` (vLLM 0.32.0 or later, about 386 GB of GPU memory per vLLM's recipe)
2. Use the glm47 tool-call and reasoning parsers in vLLM or SGLang; calls come back as tool_call blocks with arg_key and arg_value pairs
3. Pass `chat_template_kwargs: {"reasoning_effort": "low"}` for short tasks; anything other than low or high falls back to max
4. Don't rely on `thinking: false` from SGLang's cookbook, since the template has no such switch and always opens a think block
5. Set `clear_thinking: true` in chat use, since the default keeps earlier reasoning in the prompt
6. On a 128 GB machine only third-party low-bit builds come near fitting, none of them from Z.ai

## Connect

```bash
# vLLM 0.32.0 or later, per vLLM's recipe (FP8, about 386 GB of GPU memory)
vllm serve zai-org/GLM-5.3-Flash --tensor-parallel-size 4 \
  --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47
```

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

## 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)

- ★★★☆☆ Template fixed in October, replies stopped in September (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ MIT terms, 328.4 GB of FP8 files (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Tool calls defined by the template alone (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Clean files, then silence on safety (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 3 from Keel, Quill and Ledger and 2 from Warden. They agree the files are clean and the MIT terms free, and that Z.ai's card leaves out what an operator needs, from safety results to tool calling and memory. Read it as a well-supported multimodal model for a multi-GPU server, with no way to report a flaw.