GLM-5 Emerges First, Built on DeepSeek Tech, Triggering a 40% Stock Surge

From a Mysterious Codename

OpenRouter quietly introduced an anonymous model named Pony Alpha . The model’s strong performance sparked speculation that it was Zhipu AI’s next‑generation GLM‑5.0.

100% Confirmation: Pony Alpha Is GLM‑5

Users verified the model by changing the OpenRouter chat system prompt to "custom" (leaving it blank) and asking, "What model are you?" The response clearly identified the model as GLM‑5.

GitHub Clues

Several GitHub projects have already begun preparing for GLM‑5. Notably, the GLM MoE DSA implementation is slated to be added to the official transformers library.

Gigantic Scale: 745B Parameters

GLM‑5 boasts an impressive 745 billion parameters, marking it as a truly massive model.

Inheriting DeepSeek: Reusing V3 Architecture

A recent vLLM pull request (#34124) reveals that GLM‑5 builds on DeepSeek‑V3/V3.2:

Reuses the DeepSeek‑V3/V3.2 architecture

Adopts DeepSeek Sparse Attention (DSA)

Supports Multi‑Token Prediction (MTP)

This indicates that GLM‑5 follows a Mixture‑of‑Experts (MoE) design based on DeepSeek‑V3.2’s DSA roadmap.

Large‑Model Spring Festival Is Real

From Qwen 3.5 and ByteSeed to the newly identified GLM‑5, major players are launching major models during the Chinese New Year period, making 2026’s large‑model "spring recruitment" more exciting than expected.