SambaNova brings Z.ai’s GLM-5.3 to SambaCloud at over 200 tokens/s

A model aimed at long-running agent work

GLM-5.3 is Z.ai’s flagship model for complex software engineering and long-horizon agent tasks, according to SambaNova. The company says the model was trained in environments that mirror multi-day expert engineering work, including access to compute clusters, codebases, documentation and experiment results.

That training is intended to let GLM-5.3 manage a task from start to finish instead of waiting for a user to divide it into smaller steps. SambaNova says the model scores 28.3 on Terminal-Bench 3.0 and 66.9 on DeepSWE v1.1, which it describes as state-of-the-art results among open-weight models.

The company also highlights token efficiency. On Z.ai’s internal Code Bench, GLM-5.3 reached 31.4% at around 50K output tokens per task, compared with 29.5% for Claude Opus 4.8 at 120K tokens, according to SambaNova.

Speed and caching target agent loops

SambaNova’s case for running GLM-5.3 on its cloud is centered on repeated model calls. In an example trace, an agent made 41 calls over 17 minutes and 30 seconds, processed 5.4 million input tokens and generated just under 42,000 output tokens.

Because each turn resends what the agent has seen, input grew from 38K tokens on the first call to over 114K by the eighth. Output remained small, consisting of a few hundred tokens of reasoning and a tool call, while over 91% of the input was served from cache.

SambaNova says GLM-5.3 runs at over 200 tokens/s on SambaCloud. It argues that this per-call speed is important for long-running workflows, where latency across many steps can affect how long a developer waits for the full task to finish.

Access through APIs and coding tools

Developers can call the model from Python through SambaNova’s API, using the model string GLM-5.3. The example provided by SambaNova asks the model to diagnose a timeout problem in a FastAPI service, implement pagination and query timeouts, and write regression tests, with reasoning effort set to max.

GLM-5.3 is also available through ZCode, Z.ai’s agentic development environment. SambaNova says users can connect it by adding the SambaCloud OpenAI base URL, entering an API key, selecting GLM-5.3 and setting reasoning effort to max.

The same base-URL approach works with Claude Code, OpenCode and Codex CLI, although SambaNova says Codex users should use the Responses API endpoint and declare low, high and max as supported reasoning levels. The company also directs users to the SambaCloud playground and API key setup.

For developers, the announcement expands access to GLM-5.3 across tools built for coding and multi-step automation. SambaNova also points to cybersecurity use cases, saying the model holds the best published score on CyberGym and was rated by the US Center for AI Standards and Innovation as the most advanced open-weight model for cybersecurity.