SambaNova Serves GLM-5.3 at Over 200 Tokens Per Second

SambaNova Systems announced on November 11, 2026, that Z.ai's open-weight GLM-5.3 model now runs on its SambaCloud inference service at more than 200 tokens per second. Developers can reach the model through the SambaCloud playground, a standard API, and integrations for coding agents.

The announcement adds a hosted deployment for a model Z.ai first introduced on August 14, 2026. GLM-5.3 shares its base model with GLM-5.2; Z.ai attributes the newer version's gains to additional post-training across more environments, more diverse tasks, and more compute. The intended work is complex software engineering and long-horizon agent tasks built from sequences of dependent steps.

What does the agent trace reveal?

SambaNova supplied a trace from an example agent run to show what those workloads cost in practice. The run made 41 model calls over 17 minutes and 30 seconds, processing 5.4 million input tokens while generating just under 42,000 output tokens. That ratio is roughly 129 input tokens for every token produced.

Because each turn resent the accumulated context, input grew from 38,000 tokens on the first call to more than 114,000 by the eighth. Outputs stayed small — a few hundred tokens of reasoning plus a tool call — and more than 91% of the input was served from cache.

That pattern explains why serving speed dominates the pitch. Agent loops repeatedly send long contexts back to the model while requesting only a short next action, so per-call latency accumulates across the sequence. SambaNova positions its throughput as a way to keep those steps short, and the cache figure shows how heavily the workload reuses earlier context.

How strong are the reported results?

Z.ai says its post-training environments resembled professional work, including tasks that could occupy an experienced engineer for several days. In one machine-learning infrastructure scenario, the model receives access to compute clusters, storage systems, internal documentation, codebases, and experiment results. It must diagnose bottlenecks across a training stack, implement optimizations, run experiments, and deliver a measurable end-to-end speedup while preserving correctness.

The reported benchmark movement is substantial:

• Terminal-Bench 3.0: from 4.6 to 28.3
• DeepSWE v1.1: from 46.2 to 66.9
• Agents' Last Exam: from 23.8 to 28.5

Those gains map onto terminal work, software engineering, and general agent tasks — the same categories SambaNova targets with this deployment.

On its private Code Bench, Z.ai reports GLM-5.3 reaching 31.4% at High effort with about 50,000 output tokens per task, against Claude Opus 4.8 at 29.5% and 120,000 tokens. The caveat sits in the same table: Claude Fable 5 reached 39.5% at Max effort and stays ahead. GLM-5.3 generated fewer tokens and scored higher than Opus 4.8 in this test, but Z.ai's own comparison still places Fable 5 first.

What does access look like in practice?

On SambaCloud, developers call the model with the model string GLM-5.3. SambaNova's Python example sends chat-completion requests to https://api.sambanova.ai/v1 and sets reasoning_effort to max, the level Z.ai recommends for coding.

ZCode can use SambaCloud through OpenAI- or Anthropic-compatible provider settings, with a base URL, an API key, and GLM-5.3 selected in the model menu. SambaNova says the same base-URL change works with Claude Code, OpenCode, and Codex CLI. For Codex, developers must use the Responses API endpoint and declare the model's supported reasoning levels in their model catalog.

One migration detail deserves attention. GLM-5.3 supports low, high, and max reasoning effort, and thinking cannot be disabled. Z.ai says applications still using thinking.type: "disabled" must enable thinking and set reasoning effort to low before updating the model ID; otherwise, requests will fail.

The distinction that matters for buyers: benchmark scores describe what a model can complete, while the trace exposes the repeated serving work required to complete it. SambaNova's deployment targets that second half of the problem, making per-call speed and cache reuse as relevant to the user experience as GLM-5.3's reported task scores.