Today, we're introducing GLM-5.2 Fast: a separate Model API tier that serves the same GLM-5.2 weights on infrastructure tuned for per-user throughput for real-time applications. You can try it here now or switch over your GLM-5.2 standard to the endpoint zai-org/GLM-5.2-Fast.
Why agentic workloads need a different tier
When GLM-5.2 shipped, it was strong enough at coding and tool use that teams started moving real-time agentic workloads onto it.
Agentic workflows are systems: a main agent plans, delegates to subagents, reviews their work, and merges the results. Each handoff is another inference call and another place latency can compound.
GLM-5.2 Fast transforms this workflow because it’s an open model we trust as the main agent, not just for the subagents around it. It’s smart enough to coordinate the workflow, cost-effective enough to scale across subagents, and fast enough to keep the entire system competitive.
GLM-5.2 Fast preserves ease-of-use with tighter performance guardrails
GLM 5.2 Fast serves the same weights as standard GLM-5.2, on infrastructure provisioned differently for real-time agentic use cases. It offers:
Ease-of-use: Model APIs are just that; OpenAI-compatible API endpoints you point your code at. There’s no infrastructure to manage, and you pay only for what you use (per 1M tokens).
Tighter performance SLAs: our Fast tier is built to handle even the most variable, bursty workloads while maintaining a small variance in throughput and latency metrics.
Switching from standard GLM-5.2
Because Fast is just a model slug, switching is a one-line change:
1import os
2from openai import OpenAI
3
4client = OpenAI(
5 api_key=os.environ["BASETEN_API_KEY"],
6 base_url="<https://inference.baseten.co/v1>",
7)
8
9response = client.chat.completions.create(
10 model="zai-org/GLM-5.2-Fast",
11 messages=[{"role": "user", "content": "Refactor this function to be async."}],
12 stream=True,
13)You can run Fast alongside standard GLM-5.2, route by workload — agentic loops to Fast, batch summarization to standard — and see usage per tier separately in your dashboard. Most real deployments mix latency-sensitive and latency-tolerant traffic; routing each to the tier it needs cost-efficiently.
Getting started
Access is fully open at launch. Customers can access directly through the UI or through their account team. Switching over is a one-line model slug change. You can start using it here.
