GLM-4.7-Flash
Z.ai's MIT-licensed 30B-A3B model for local coding and agents, from January 2026 and still downloaded about a million times a month.
In short. GLM-4.7-Flash is a small open model from Z.ai (Zhipu) for coding and agents. It has about 30 billion parameters but uses only 3 billion for each token, so it is fast. The MIT license has almost no conditions, and by our estimate a 4-bit build needs about 22 GB, so it fits a gaming GPU or a laptop with 32 GB.
Summary
Z.ai released GLM-4.7-Flash on 19 January 2026 under the MIT license and called it the strongest model of the 30B class, a lightweight option next to the full GLM-4.7. It is a 30B-A3B mixture of experts, with 31.2B parameters on Hugging Face: 47 layers, 64 routed experts with 4 chosen per token plus one shared expert, and a configuration that lists 202,752 positions. It reads and writes text. Z.ai says it does well in its internal frontend and backend coding tests and also recommends it for Chinese writing, translation, long context and role play. For multi-turn agent tasks the card asks you to turn on its Preserved Thinking mode. None of the GLM-5 releases that followed is small: the smallest, GLM-5.3-Flash, has 320B parameters.
In Z.ai's own table, GLM-4.7-Flash is compared with Qwen3-30B-A3B-Thinking-2507 and gpt-oss-20b. It scores 91.6 on AIME 25 (gpt-oss-20b 91.7, Qwen 85.0), 75.2 on GPQA (73.4 and 71.5), 64.0 on LiveCodeBench v6 (Qwen 66.0, gpt-oss-20b 61.0), 14.4 on HLE, 59.2 on SWE-bench Verified (22.0 and 34.0), 79.5 on τ²-Bench (49.0 and 47.7) and 42.8 on BrowseComp (2.29 and 28.3). All of these are Z.ai's own results, and the models it compares against are from 2025. Newer open models in the same size class, such as Qwen3.5 in February and Qwen3.6 in April 2026, came later.
The only official weights are BF16, 62.5 GB, so most people run community quantizations. By our estimate a 4-bit build needs about 22 GB. At launch the card said vLLM and SGLang supported the model only on their main branches, and it also loads in Transformers. In the 30 days to 11 November 2026, Hugging Face counted about 1,075,000 downloads of the official repository and about 302,000 of Unsloth's GGUF repository, though download counts can include automated pulls.
Good for
- Local coding and chat on a 24 GB GPU or a 32 GB Mac, using a community 4-bit build
- Tool-calling agents that need a small model with the permissive MIT license
- Chinese and English writing and translation
- Fine-tuning and experiments on a single machine
Watch out for
- All scores are Z.ai's own and compare against 2025 models, and Qwen3.5 and Qwen3.6 models of similar size came later
- Only a 62.5 GB BF16 version is official, so local use relies on community quantizations of varying quality
- At launch vLLM and SGLang needed main-branch builds, so check the current releases
- It reads text only, and the card lists English and Chinese
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the GLM-4.7-Flash (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| GLM-4.7-Flash (BF16) BF16 | 31.2B 3B active | 72 GB | 36 GB | 22 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is GLM-4.7-Flash?
- GLM-4.7-Flash is a small open model from Z.ai (Zhipu) for coding and agents. It has about 30 billion parameters but uses only 3 billion for each token, so it is fast. The MIT license has almost no conditions, and by our estimate a 4-bit build needs about 22 GB, so it fits a gaming GPU or a laptop with 32 GB.
- Can I use GLM-4.7-Flash commercially?
- Yes. GLM-4.7-Flash is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is GLM-4.7-Flash open source?
- GLM-4.7-Flash is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run GLM-4.7-Flash on my own computer?
- Yes, on a good graphics card. The smallest variant, GLM-4.7-Flash (BF16), needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was GLM-4.7-Flash released?
- GLM-4.7-Flash was released by Z.ai (Zhipu AI) on January 19, 2026.
- How long a context window does GLM-4.7-Flash have?
- GLM-4.7-Flash supports a context window of up to 198K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- GLM-4.7-Flash, open source and free (Z.ai news, 19 January 2026) zhipuai.cn
- GLM-4.7-Flash model card (Hugging Face) huggingface.co
- GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models (the technical report the card links) arxiv.org
- Z.ai release notes (GLM-4.7 and later releases) docs.z.ai
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| GLM-5.3-Flash Z.ai (Zhipu AI) | 320B 18B active | 1M | MIT (Permissive) | 221 GB Big-memory desktop | |
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone |