GLM-4.7-Flash Explained: Zhipu's 30B MoE Model, Benchmarks, Local Deployment and Free API

GLM-4.7-Flash is an open-weight model from Zhipu AI (Z.ai) that targets the “30B class”: large enough to handle real coding and agent tasks, small enough to deploy on a single high-end workstation GPU. It is a mixture-of-experts (MoE) model with about 30 billion total parameters and about 3 billion active per token, released under the MIT licence. This article covers what the model is, what the published benchmarks say, how to run it locally or via API, and what to keep in mind before adopting it.

What GLM-4.7-Flash is

The official model card describes it as a “30B-A3B MoE model” and positions it as the strongest model in its size class, intended as a lightweight deployment option that balances performance and efficiency. It is the small sibling of GLM-4.7, Z.ai’s flagship model, described in the GLM-4.7 technical blog.

Why MoE matters here

In an MoE model, a router sends each token to a small subset of “experts”. Only around 3B parameters are active per token, so compute per token is closer to a 3B dense model, while the full 30B parameters provide more capacity. The catch is memory: all weights still need to be loaded, so VRAM or RAM requirements follow the total parameter count, not the active one. Quantised builds reduce this, at some cost in quality.

Published benchmarks

Z.ai’s model card compares GLM-4.7-Flash with two similarly sized open models:

Benchmark	GLM-4.7-Flash	Qwen3-30B-A3B-Thinking-2507	GPT-OSS-20B
AIME 25	91.6	85.0	91.7
GPQA	75.2	73.4	71.5
LiveCodeBench v6	64.0	66.0	61.0
HLE	14.4	9.8	10.9
SWE-bench Verified	59.2	22.0	34.0
τ²-Bench	79.5	49.0	47.7
BrowseComp	42.8	2.29	28.3

Two things stand out. First, the biggest gaps are on agentic benchmarks (SWE-bench Verified, τ²-Bench, BrowseComp), which test multi-step tool use rather than single answers. Second, on pure reasoning and coding puzzles (AIME 25, LiveCodeBench), the three models are close, and GLM-4.7-Flash is not ahead on every row.

These are vendor-reported numbers. The card lists its evaluation settings, for example temperature 1.0 and top-p 0.95 for most tasks, and “Preserved Thinking” mode for multi-turn agent tasks. For τ²-Bench, it notes extra prompts and domain fixes were applied. Results in your own harness may differ, so treat the table as a starting point.

Running it locally

The model card lists vLLM, SGLang and Hugging Face transformers. At the time of writing, vLLM and SGLang support required their main or nightly branches.

vLLM, per the card:

pip install -U vllm --pre --index-url https://pypi.org/simple \
  --extra-index-url https://wheels.vllm.ai/nightly
pip install git+https://github.com/huggingface/transformers.git

vllm serve zai-org/GLM-4.7-Flash --tensor-parallel-size 1

transformers works for quick experiments by loading zai-org/GLM-4.7-Flash with AutoModelForCausalLM and applying the chat template.

Community-quantised GGUF builds also exist for llama.cpp-based tools such as Ollama and LM Studio. These are not official releases, so check who published a build and which quantisation level it uses.

Hardware considerations

Full-precision 30B weights need far more memory than a typical 24 GB consumer card provides, so single-GPU local use usually means a quantised build. Long contexts also add KV-cache memory on top of the weights. If you plan to use long documents, test memory use at the context length you actually need.

Using it through the API

If local hardware is a constraint, Z.ai offers the model through its API. Its pricing page lists GLM-4.7-Flash as free for input, cached input and output, while a faster GLM-4.7-FlashX variant is priced at $0.07 per million input tokens and $0.40 per million output tokens. Free access is subject to account rate and concurrency limits. The API is OpenAI-style, so many existing tools can connect by changing the base URL and model name.

Where it fits

• Local coding assistants and agents, where the SWE-bench and τ²-Bench results suggest stronger tool use than peers of similar size.
• Cost-sensitive pipelines, using the free API tier for prototyping and moving to self-hosting when volume grows.
• Privacy-sensitive work, where data must stay on your own hardware.

Comparison with alternatives

• Qwen3-30B-A3B has the same 30B/3B MoE shape and a large ecosystem; it leads slightly on LiveCodeBench in the table above.
• GPT-OSS-20B from OpenAI is smaller and matches GLM-4.7-Flash on AIME 25, but trails on the agentic benchmarks listed.
• Larger models such as the full GLM-4.7 remain stronger on hard tasks; Flash trades some capability for deployability.

Limitations

• Benchmarks are self-reported and some used custom prompts or fixes.
• Early inference support depended on development branches, which can change.
• Quantised community builds vary in quality.
• Like any LLM, it can produce incorrect code; review and test generated output.

Sources

• GLM-4.7-Flash model card (Hugging Face)
• GLM-4.7 technical blog (Z.ai)
• Z.ai API pricing
• GLM-4.5 technical report (arXiv 2508.06471)