---

license: mit

language:

• en
• zh

base_model:

• zai-org/GLM-4.7-Flash

tags:

• large-language-model
• mixture-of-experts
• reasoning
• agentic-ai
• agentic-coding
• code-generation
• tool-use
• function-calling
• long-context
• multilingual
• gguf

---

GLM-4.7-Flash

GLM-4.7-Flash is a lightweight Mixture-of-Experts language model developed by Z.ai, designed to provide a strong combination of reasoning, coding, agentic task execution, and inference efficiency. It contains approximately 30 billion total parameters while activating only around 3 billion parameters per token, allowing it to provide substantially lower active computation than a dense model of comparable total parameter count.

This repository contains GGUF quantized variants of GLM-4.7-Flash optimized for efficient local inference. The quantized formats substantially reduce the model's storage and memory requirements while retaining its core reasoning, coding, and tool-use capabilities.

GLM-4.7-Flash is particularly positioned for agentic coding and multi-step reasoning rather than being a simple conversational language model. It supports thinking modes, function calling, structured output, streaming, context caching, and long-context interactions. Z.ai reports strong results across mathematical reasoning, general reasoning, software engineering, and agent benchmarks.

---

Model Overview

• Model Name: GLM-4.7-Flash
• Base Model: zai-org/GLM-4.7-Flash
• Architecture: Mixture-of-Experts (MoE) Transformer
• Total Parameters: Approximately 30 Billion
• Active Parameters: Approximately 3 Billion per token
• Attention: Multi-head Latent Attention (MLA)
• Modalities: Text
• Languages: English, Chinese
• Maximum Context: Up to approximately 200K tokens
• Maximum Output: Up to 128K tokens
• Developer: Z.ai / Zhipu AI
• License: MIT

---

Quantization Formats

This repository provides various GGUF quantized versions of GLM-4.7-Flash optimized for efficient local inference.

IQ3_M

• Size reduction of approx 77.95% (12.30 GB) compared to 16-bit (55.79 GB)
• Aggressive 3-bit quantization designed to substantially reduce the memory footprint of the 30B-class MoE model
• Suitable for resource-conscious local deployments where reducing model size is a primary requirement
• Retains the model's sparse MoE architecture while making deployment considerably more accessible than the F16 version
• Complex reasoning, long-horizon agentic coding, and extended tool-use workflows may experience greater quality degradation than higher-precision variants

IQ4_NL

• Size reduction of approx 71.52% (15.89 GB) compared to 16-bit (55.79 GB)
• Advanced 4-bit non-linear quantization designed to preserve reasoning and generation quality while substantially reducing storage requirements
• Well suited for coding, mathematical reasoning, agentic workflows, structured generation, and tool-assisted applications
• Provides a stronger fidelity-to-size trade-off than the more aggressive IQ3_M format
• Requires more memory than IQ3_M while retaining a larger portion of the original model's generation behavior

IQ4_XS

• Size reduction of approx 73.02% (15.05 GB) compared to 16-bit (55.79 GB)
• Balanced 4-bit quantization focused on efficient deployment without heavily compromising reasoning and coding behavior
• Suitable for local coding agents, long-context assistants, tool-calling systems, and general reasoning workloads
• Offers a practical compromise between model footprint, generation quality, and inference efficiency
• Provides a smaller footprint than IQ4_NL while remaining substantially less aggressive than IQ3_M

---

Model Architecture

GLM-4.7-Flash uses a Mixture-of-Experts architecture rather than a conventional dense Transformer. The model contains 64 routed experts and activates 4 routed experts for each token, together with a shared expert. Consequently, although the model contains roughly 30B total parameters, only a fraction of those parameters participate in the computation for each token.

The configuration contains 47 hidden layers, a hidden size of 2,048, 20 attention heads, and 20 key/value heads. It also uses the GLM4 MoE Lite architecture with latent attention components designed for efficient processing.

This sparse architecture is a major reason GLM-4.7-Flash can provide capabilities associated with a much larger model while keeping the active computation considerably lower.

---

Reasoning and Agentic Capabilities

GLM-4.7-Flash is designed around task completion rather than isolated response generation. Z.ai specifically highlights improvements in programming, multi-step reasoning, tool collaboration, and complex agentic execution.

Thinking

The model supports multiple thinking modes for adapting the amount of reasoning to the task. Preserved Thinking is particularly useful for multi-turn agentic workloads because reasoning information can be retained across interactions.

Tool Use

The model supports function calling and can be integrated with external tools and services. This makes it suitable for agents that need to retrieve information, execute functions, manipulate data, or interact with external systems.

Long-Context Processing

The model supports a context length of approximately 200K tokens, with up to 128K output tokens according to Z.ai's model documentation. This enables large codebase analysis, extended conversations, long documents, and multi-step agent sessions.

---

Key Capabilities

• Agentic Coding

Designed to complete software-development tasks through planning, reasoning, code generation, and tool interaction rather than only generating isolated code snippets.

• Mathematical Reasoning

Demonstrates strong performance on mathematical reasoning benchmarks, including AIME 25.

• Software Engineering

Supports code generation, debugging, repository-level reasoning, and multi-step development workflows.

• Tool Calling

Provides function-calling capabilities for integrating external tools and services into agent pipelines.

• Structured Output

Supports structured formats such as JSON for integration with downstream applications.

• Long-Context Reasoning

Handles very large contexts suitable for extended conversations, codebases, documents, and agent trajectories.

• Multilingual Text Processing

The official model card identifies English and Chinese support.

Recommended Usecases

• Agentic Coding

Build coding agents capable of planning, implementing, debugging, and iterating over software tasks.

• Software Engineering Assistants

Support repository analysis, code generation, debugging, and development workflows.

• Tool-Calling Agents

Connect the model with APIs, databases, search systems, terminal tools, and other external functions.

• Long-Context Applications

Process large documents, codebases, conversations, and extended agent histories.

• RAG Systems

Build retrieval-augmented assistants capable of reasoning over large retrieved contexts.

• Mathematical & Analytical Workloads

Apply the model to complex reasoning and structured analytical tasks.

• Enterprise AI Agents

Develop internal assistants capable of combining reasoning, structured output, and external tool execution.

---

Usage Example

Using llama.cpp

./llama-cli \
  -m SandLogicTechnologies/GLM-4.7-Flash_IQ4_NL.gguf \
  -p "Analyze this software architecture and propose a step-by-step implementation plan."

For long-context and agentic workloads, the available context size should be configured according to the available system memory and inference backend.

Acknowledgments

These quantized models are based on the original work by the Z.ai / GLM Team.

Special thanks to:

• The Z.ai / GLM Team for developing and releasing GLM-4.7-Flash.
• The developers and open-source community behind llama.cpp for enabling efficient GGUF-based local inference.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/