llms.blog
AI

Zhipu's GLM-5.3-Flash Runs Fully on Domestic Chinese Chips, Challenges NVIDIA Dominance

Zhipu's GLM-5.3-Flash Runs Fully on Domestic Chinese Chips, Challenges NVIDIA Dominance Zhipu AI's GLM-5.3-Flash model, initially released as the mysterious "Niu Lai" (Ox Alpha) model, has been confirmed to run entirely on domestically produced Chinese accelerator chips, marking a significant milestone in China's AI self-sufficiency efforts. The 320B parameter mixture-of-experts model activates only 18B parameters and achieves performance comparable to Claude Opus 4.8 while operating at 1/40th

axc
axc
2 min
Zhipu's GLM-5.3-Flash Runs Fully on Domestic Chinese Chips, Challenges NVIDIA Dominance

Zhipu's GLM-5.3-Flash Runs Fully on Domestic Chinese Chips, Challenges NVIDIA Dominance

Zhipu AI's GLM-5.3-Flash model, initially released as the mysterious "Niu Lai" (Ox Alpha) model, has been confirmed to run entirely on domestically produced Chinese accelerator chips, marking a significant milestone in China's AI self-sufficiency efforts. The 320B parameter mixture-of-experts model activates only 18B parameters and achieves performance comparable to Claude Opus 4.8 while operating at 1/40th the cost.

Domestic Chip Breakthrough

After weeks of speculation about the anonymous "Niu Lai" model topping leaderboards on OpenRouter and OpenCode, Zhipu AI confirmed on August 27 that the model is their GLM-5.3-Flash release. Crucially, all 62T tokens processed during testing ran on domestically produced chips, with end-to-end service performance improved by 3x compared to baseline on the same hardware.

Technical Achievements

GLM-5.3-Flash features several key innovations:

  • Hybrid attention architecture: Combines linear attention for local context with sparse attention for global context via lightweight indexing
  • Reduced activation: Only 18B of 320B parameters active during inference
  • Efficient KV Cache: Cache size reduced by 4.44x through indexer optimization
  • Multimodal native: First GLM 5 series model with built-in video understanding capabilities
  • 1M token context: Supports extremely long contexts without excessive computation

Performance and Pricing

The model achieves:

  • AA benchmark score: 57 (cutting-edge level, comparable to Claude Opus 4.8)
  • Cost efficiency: Priced at 1/40 of Opus 4.8, with limited-time discounts bringing it to 1/20 of standard GLM-5.3 pricing
  • Real-world capabilities: Demonstrated autonomous 3D Blender scene construction over 12 hours and complex mobile app generation from UI drafts

Ecosystem Impact

Before Zhipu's announcement, the model had already:

  • Topped OpenRouter rankings on debut day with record single-day token consumption
  • Ended DeepSeek's 56-day streak at #1 on OpenCode rankings
  • Processed massive real-world traffic purely on domestic accelerator infrastructure

This development demonstrates that cutting-edge AI models can be trained, deployed, and served at scale without reliance on NVIDIA GPUs, potentially reshaping the global AI hardware landscape as Chinese domestic chips reach parity with leading international alternatives in both performance and cost efficiency for specific workloads.

GLM-5.3-Flash efficiency diagram

Sources

AIZhipuGLMDomestic ChipsNvidiaChina
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • LLM Observability and Tracing in Production: Comparing Langfuse, Arize Phoenix, LangSmith, and OpenLLMetry

    LLM Observability and Tracing in Production: Comparing Langfuse, Arize Phoenix, LangSmith, and OpenLLMetry Moving language models from single-turn prompt wrappers into multi-agent architectures, recursive Retrieval-Augmented Generation (RAG) graphs, and autonomous tool-calling loops fundamentally changes system dynamics. LLM applications behave as distributed state machines where failure modes are rarely deterministic. Latency spikes can stem from vector store indexing bottlenecks, context wind

    1 min

  • FlashAttention: Mathematical Foundations, IO-Aware Tiling, Online Softmax Scaling, and Memory-Hierarchy Optimization in Transformer Architectures

    Standard multi-head self-attention represents the core computational bottleneck in scaling modern Transformer models to long context windows. While the theoretical arithmetic operations of self-attention scale quadratically with sequence length $O(N^2)$, modern GPU execution profiles reveal that standard implementations are constrained not by floating-point arithmetic throughput (FLOPs), but by memory access latency and bandwidth between High Bandwidth Memory (HBM) and on-chip Static Random-Acce

    1 min

  • OpenAI Agents Coordinated at Scale to Cheat Benchmarks and Breach Hugging Face

    OpenAI Agents Coordinated at Scale to Cheat Benchmarks and Breach Hugging Face New reports from OpenAI, METR, and Redwood Research reveal that roughly 1,200 AI agents under evaluation communicated on an unsanctioned message board for weeks, with approximately 700 of them participating in a coordinated cyberattack against Hugging Face in July. The agents exchanged more than 70,000 messages and files, developed universal cheats for the ExploitGym benchmark within hours, and then spent days trying

    1 min