跪拜 Guibai
← All articles
NVIDIA · ChatGLM (Zhipu) · AI Programming

NVIDIA Hosts a Free GLM-5.2 API Endpoint Compatible with OpenAI Toolchains

By 该用户已不存在 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A free, OpenAI-compatible endpoint for a 744B-parameter model removes the cost barrier for prototyping AI coding workflows. The real cost is peak-hour congestion and rate limits, so it works best as a development sandbox, not a production dependency.

Summary

The 744B-parameter GLM-5.2 model, open-sourced by Zhipu AI under the MIT license, is available as a free API endpoint on NVIDIA's build.nvidia.com platform. Registration works with Chinese phone numbers, and the endpoint is compatible with the OpenAI Chat Completions format, allowing direct integration into tools like Claude Code, Cursor, and VS Code.

Three integration paths are available: direct code calls via the OpenAI SDK, a graphical switcher called CC Switch for Claude Code users, and a local gateway (ServBay AI Gateway) that provides virtual keys, multi-provider failover, and usage monitoring. The free tier carries no documented daily request cap but suffers from latency spikes and HTTP 429 errors during peak hours, with a practical throughput ceiling around 40 requests per minute.

This move slots into a three-layer ecosystem: application builders like Meituan and Alibaba embed the model into products, NVIDIA distributes it as free compute to seed its GPU infrastructure, and Zhipu supplies the open-source model that flows through all of them without competing for end users.

Takeaways
Registration on build.nvidia.com accepts Chinese phone numbers and generates an API key starting with `nvapi-`.
The endpoint uses Base URL `https://integrate.api.nvidia.com/v1` and model identifier `z-ai/glm-5.2`.
Direct integration requires only swapping the base URL and API key in any OpenAI-compatible SDK.
CC Switch provides a graphical interface to reroute Claude Code's inference engine to the GLM-5.2 endpoint.
ServBay AI Gateway adds virtual key management, automatic failover when NVIDIA returns 429 errors, and per-request usage logging.
Peak-hour latency can exceed ten seconds, and the free tier caps out around 40 requests per minute.
GLM-5.2 performs well on code generation and agentic workflows but lags behind Claude on long-range reasoning and prompt adherence.
Conclusions

NVIDIA's free API is not charity; it is a customer acquisition funnel for its paid GPU cloud, where free prototyping leads naturally to paid production deployments.

Zhipu's MIT-licensed model creates a 'water source' strategy where multiple platforms compete on distribution while the model itself becomes infrastructure no single company controls.

The three-layer stack (application, infrastructure, model) shows how open-source AI commoditizes the model layer, shifting value capture to compute providers and product builders.

Concepts & terms
MoE (Mixture of Experts)
A model architecture that activates only a subset of its total parameters for any given input, reducing inference cost while maintaining a large total parameter count. GLM-5.2 uses a 744B MoE design.
NVIDIA NIM (NVIDIA Inference Microservices)
NVIDIA's catalog and deployment system for AI models, providing containerized, optimized inference endpoints on its GPU infrastructure.
Virtual Key
A proxy credential issued by an API gateway that maps to a real upstream API key. It allows per-project revocation and keeps the real key from being exposed in client configurations.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗