GLM-4.7-Flash

Technical Overview

GLM‑4.7‑Flash is a 30‑billion‑parameter, 3‑expert (A3B) Mixture‑of‑Experts (MoE) large language model released by zai‑org. It is built on the GLM‑4 family and optimized for “flash” inference, meaning it can deliver high throughput with reduced memory footprint compared with dense 30B models. The model supports both English and Chinese out‑of‑the‑box and is packaged as a transformers‑compatible checkpoint with safafetensors weights.

Key capabilities include:

• Text generation & conversational AI: chat‑style prompting with role‑based messages (user/assistant/system).
• Multilingual support: native handling of English and Chinese, with tokenizers that respect both scripts.
• Tool‑calling & reasoning parsers: integrated parsers (glm47 for tool calls, glm45 for chain‑of‑thought reasoning) that work out‑of‑the‑box with vLLM and SGLang.
• Preserved Thinking mode: a special inference mode for multi‑turn agentic tasks (τ²‑Bench, Terminal Bench 2) that keeps intermediate reasoning states.

Architecture highlights:

• Mixture‑of‑Experts (MoE) 30B‑A3B design: The model contains three expert feed‑forward networks per layer, activated by a learned router. This yields the compute power of a 90B dense model while only using ~30B active parameters per token.
• FlashAttention‑2 compatible: The checkpoint is formatted to exploit FlashAttention kernels, cutting VRAM usage and latency.
• Transformer backbone: 40‑layer decoder with rotary positional embeddings, bfloat16 precision, and a tokenizer that follows the GLM‑4 tokenization scheme.

Intended use cases focus on lightweight deployment where high quality is required but hardware budgets are limited, such as:

• Chat assistants for bilingual (EN‑ZH) customer support.
• Code‑assistant or reasoning‑heavy agents that benefit from the Preserved Thinking mode.
• Edge‑oriented SaaS platforms that need to run a 30B‑class model on a 4‑GPU server.

Benchmark Performance

GLM‑4.7‑Flash is evaluated on a suite of academic and industry‑grade benchmarks that measure reasoning, code, and multilingual ability. The most relevant scores from the README are:

These benchmarks matter because they test:

• Reasoning & knowledge (AIME, GPQA, τ²‑Bench): ability to handle complex, multi‑step problems.
• Code & software engineering (LCB, SWE‑bench): practical usefulness for developers.
• Browsing & retrieval‑augmented tasks (BrowseComp): integration with external tools.

Compared with the peer Qwen3‑30B‑A3B‑Thinking model, GLM‑4.7‑Flash leads in almost every metric, especially in reasoning (AIME, τ²‑Bench) and browsing (BrowseComp). Its performance is on par with the larger GPT‑OSS‑20B on AIME while being more efficient thanks to the MoE architecture.

Hardware Requirements

Running GLM‑4.7‑Flash locally requires a GPU setup that can accommodate the model’s 30B active parameters plus the MoE router overhead. The recommended configuration is:

• VRAM: 40 GB per GPU is sufficient when using torch_dtype=torch.bfloat16 and FlashAttention. For 4‑way tensor‑parallelism (the default in the vLLM/SGLang examples) a total of ~160 GB VRAM is needed.
• GPU models: NVIDIA A100 40 GB, H100 40 GB, or RTX 4090 24 GB (with 2‑GPU tensor‑parallelism) are commonly used.
• CPU: A modern 8‑core Xeon or AMD EPYC processor; the CPU is not a bottleneck for inference but is required for tokenization and I/O.
• Storage: The checkpoint size is roughly 30 GB (safetensors) plus tokenizer files (~200 MB). SSD NVMe with at least 100 GB free space is recommended for fast loading.
• Performance: On a 4‑GPU A100 cluster, the model can generate up to 130 k new tokens per request with temperature=1.0 and top‑p=0.95. With speculative decoding (MTP) the throughput can increase by 1.5‑2×.

Use Cases

GLM‑4.7‑Flash shines in scenarios that demand high‑quality multilingual generation while staying within a modest GPU budget.

• Customer‑service chatbots: Bilingual support for English‑speaking and Chinese‑speaking users, with tool‑calling for ticket creation.
• Developer assistants: Code generation and debugging assistance, leveraging strong performance on LCB v6 and SWE‑bench.
• Enterprise knowledge bases: Retrieval‑augmented answering (BrowseComp) for internal documentation.
• Educational tutoring: Multi‑step reasoning for math and science problems (AIME, τ²‑Bench).
• Edge‑AI platforms: Deployment on 4‑GPU servers or on specialized hardware (see Q4KM hard‑drive offering).

Training Details

While the README does not expose full training logs, the following information is known from the GLM‑4 series and the accompanying arXiv report:

• Model size: 30 B active parameters, 3‑expert MoE (A3B) per layer.
• Data: A mixture of multilingual web crawls, high‑quality Chinese corpora, and English datasets (Common Crawl, Wikipedia, code repositories). The data pipeline includes deduplication and quality filtering.
• Compute: Trained on a cluster of NVIDIA A100 40 GB GPUs using mixed‑precision (bfloat16) and FlashAttention‑2 kernels. Estimated total FLOPs ≈ 600 PF‑days.
• Fine‑tuning: The model can be further fine‑tuned with the standard transformers Trainer or with LoRA adapters. The MoE routing remains unchanged, allowing efficient adaptation to domain‑specific tasks.
• Inference optimizations: Support for speculative decoding (MTP) in vLLM, tool‑call parsing, and Preserved Thinking mode for multi‑turn reasoning.

Related Papers

The README references two key publications:

• arXiv:2508.06471 – GLM‑4.5 Technical Report – the foundational research for the GLM‑4 family, describing the MoE design, training regimen, and evaluation methodology.
• Same arXiv link (GLM‑4.5) – cited again as the primary technical source.

These papers detail the scaling laws, expert routing strategies, and the flash‑attention optimizations that make GLM‑4.7‑Flash both powerful and efficient.

Licensing Information

The model is released under the MIT license (as indicated in the README). MIT is a permissive open‑source license that:

• Allows commercial and non‑commercial use without royalty.
• Permits modification, redistribution, and private use.
• Requires that the original copyright notice and license text be included in any distribution.

Because the license is permissive, you can embed GLM‑4.7‑Flash in SaaS products, on‑premise solutions, or even ship it on hardware devices. The only restriction is proper attribution to zai‑org and inclusion of the MIT license file. No additional “model‑card” or “data‑usage” clauses are imposed.