README.md

---
pipeline_tag: text-generation
license: mit
base_model: zai-org/GLM-4.7-Flash
library_name: zeromodels
language:
- en
- zh
tags:
- keras
- zeromodels
- glm
- glm4_moe_lite
- glm-4.7
- mixture-of-experts
- text-generation
- pytorch
- jax
- tf
---

# Run GLM-4.7-Flash with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-181717?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-GLM-1f6feb)](https://imvision12.github.io/ZeroModels/glm4_moe_lite/) [![HuggingFace](https://img.shields.io/badge/HuggingFace-GLM-ffd21e?logo=huggingface&logoColor=black)](https://huggingface.co/collections/zeromodels/glm-6a82b8f9f753e8dcae3ff3f7)

# zeromodels/glm-4.7-flash

Pure-**Keras 3** conversion of [`zai-org/GLM-4.7-Flash`](https://huggingface.co/zai-org/GLM-4.7-Flash) for [zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on **TensorFlow / Torch / JAX**. **GLM-4.7-Flash** is a mixture-of-experts model (MLA + DeepSeekMoE) served as **text -> text**; weights are stored in **bfloat16**, with the mixture-of-experts router correction bias kept in **float32** (matching the upstream mixed-precision checkpoint). See `zm_config.json` (`weight_dtype` + `weight_dtype_overrides`) for the exact layout.

For model details, license, and usage terms, see the upstream [model card](https://huggingface.co/zai-org/GLM-4.7-Flash).

Paper: [ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools (arXiv:2406.12793)](https://arxiv.org/abs/2406.12793) · [HF Papers](https://huggingface.co/papers/2406.12793)

## ✨ Quick start

```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.glm4_moe_lite import Glm4MoeLiteTextGenerate, Glm4MoeLiteTokenizer

model = Glm4MoeLiteTextGenerate.from_weights("zeromodels/glm-4.7-flash")
tokenizer = Glm4MoeLiteTokenizer.from_weights("zeromodels/glm-4.7-flash")

messages = [{"role": "user", "content": "Name three prime numbers."}]
inputs = tokenizer(messages)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0]))
```

Load any GLM variant the same way with `from_weights("zeromodels/<variant>")`. Browse them all in the [GLM collection](https://huggingface.co/collections/zeromodels/glm-6a82b8f9f753e8dcae3ff3f7).

## Special Thanks

A huge thank you to the Zhipu AI / THUDM team for creating and releasing the GLM models.

License: `mit` (per the upstream model card).