GLM-4.7-Flash is a brand new member of the GLM 4.7 household and targets builders who need sturdy coding and reasoning efficiency in a mannequin that’s sensible to run regionally. Zhipu AI (Z.ai) describes GLM-4.7-Flash as a 30B-A3B MoE mannequin and presents it because the strongest mannequin within the 30B class, designed for light-weight deployment the place efficiency and effectivity each matter.
Mannequin class and place contained in the GLM 4.7 household
GLM-4.7-Flash is a textual content era mannequin with 31B params, BF16 and F32 tensor varieties, and the structure tag glm4_moe_lite. It helps English and Chinese language, and it’s configured for conversational use. GLM-4.7-Flash sits within the GLM-4.7 assortment subsequent to the bigger GLM-4.7 and GLM-4.7-FP8 fashions.
Z.ai positions GLM-4.7-Flash as a free tier and light-weight deployment choice relative to the complete GLM-4.7 mannequin, whereas nonetheless concentrating on coding, reasoning, and normal textual content era duties. This makes it fascinating for builders who can not deploy a 358B class mannequin however nonetheless need a trendy MoE design and robust benchmark outcomes.
Structure and context size
In a Combination of Specialists structure of this kind, the mannequin shops extra parameters than it prompts for every token. That enables specialization throughout consultants whereas retaining the efficient compute per token nearer to a smaller dense mannequin.
GLM 4.7 Flash helps a context size of 128k tokens and achieves sturdy efficiency on coding benchmarks amongst fashions of comparable scale. This context measurement is appropriate for giant codebases, multi-file repositories, and lengthy technical paperwork, the place many present fashions would want aggressive chunking.
GLM-4.7-Flash makes use of an ordinary causal language modeling interface and a chat template, which permits integration into present LLM stacks with minimal adjustments.
Benchmark efficiency within the 30B class
The Z.ai workforce compares GLM-4.7-Flash with Qwen3-30B-A3B-Considering-2507 and GPT-OSS-20B. GLM-4.7-Flash leads or is aggressive throughout a mixture of math, reasoning, lengthy horizon, and coding agent benchmarks.
This above desk showcase why GLM-4.7-Flash is among the strongest mannequin within the 30B class, a minimum of among the many fashions included on this comparability. The vital level is that GLM-4.7-Flash shouldn’t be solely a compact deployment of GLM but in addition a excessive performing mannequin on established coding and agent benchmarks.
Analysis parameters and pondering mode
For many duties, the default settings are: temperature 1.0, high p 0.95, and max new tokens 131072. This defines a comparatively open sampling regime with a big era finances.
For Terminal Bench and SWE-bench Verified, the configuration makes use of temperature 0.7, high p 1.0, and max new tokens 16384. For τ²-Bench, the configuration makes use of temperature 0 and max new tokens 16,384. These stricter settings cut back randomness for duties that want secure device use and multi step interplay.
Z.ai workforce additionally recommends turning on Preserved Considering mode for multi flip agentic duties equivalent to τ²-Bench and Terminal Bench 2. This mode preserves inside reasoning traces throughout turns. That’s helpful if you construct brokers that want lengthy chains of operate calls and corrections.
How GLM-4.7-Flash matches developer workflows
GLM-4.7-Flash combines a number of properties which might be related for agentic, coding targeted purposes:
- A 30B-A3B MoE structure with 31B params and a 128k token context size.
- Robust benchmark outcomes on AIME 25, GPQA, SWE-bench Verified, τ²-Bench, and BrowseComp in comparison with different fashions in the identical desk.
- Documented analysis parameters and a Preserved Considering mode for multi flip agent duties.
- First-class assist for vLLM, SGLang, and Transformers primarily based inference, with prepared to make use of instructions.
- A rising set of finetunes and quantizations, together with MLX conversions, within the Hugging Face ecosystem.
Take a look at the Model weight. Additionally, be happy to comply with us on Twitter and don’t neglect to affix our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Michal Sutter is an information science skilled with a Grasp of Science in Knowledge Science from the College of Padova. With a stable basis in statistical evaluation, machine studying, and knowledge engineering, Michal excels at remodeling complicated datasets into actionable insights.
