SambaNova Brings GLM-5.3 to SambaCloud at Over 200 Tokens Per Second
What happened
On November 11, 2026, SambaNova Systems announced that Z.ai’s GLM-5.3 model is now available on the SambaCloud inference service. The model can be accessed through the SambaCloud playground, API, and integrations for coding agents. This deployment is designed to handle complex software engineering and agent tasks that require processing large amounts of contextual information. The system delivers over 200 tokens per second and uses caching to efficiently serve repeated context in agent workflows.
Key details
- GLM-5.3 was introduced by Z.ai on August 14, 2026, using the same base model as GLM-5.2 with additional post-training.
- A sample agent run processed 5.4 million input tokens and generated just under 42,000 output tokens over 17.5 minutes.
- Over 91% of input tokens in the agent example were served from cache, demonstrating heavy reuse of context.
- GLM-5.3 achieved 31.4% at High effort on Z.ai’s private Code Bench, outperforming Claude Opus 4.8 but trailing Claude Fable 5.
- Developers can access GLM-5.3 on SambaCloud via API, playground, and integrations with tools like ZCode and Codex CLI.
Why it matters
Software engineering teams and developers using agent-based workflows can now run GLM-5.3 on SambaCloud with high throughput and efficient handling of large context windows. This reduces the time required for each step in long, multi-turn agent tasks, potentially lowering costs and improving productivity for organizations that depend on automated coding and reasoning tools. Teams evaluating hosted agent inference now have a deployment focused on per-call speed and cache reuse, which are critical for complex, context-heavy workloads.