SambaNova Hosts Z.ai's 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 it three ways: the SambaCloud playground, a standard API, and integrations built for coding agents.

The model itself isn't new. Z.ai introduced GLM-5.3 on August 14, 2026, and it uses the same base model as GLM-5.2. The performance jump, according to Z.ai, comes from additional post-training — more environments, more diverse tasks, and more compute. What SambaNova adds is a hosted deployment tuned for the way this class of model actually gets used: complex software engineering and long-horizon agent tasks that chain dependent steps together.

That's the real story here, and it's worth unpacking. A model can score well on benchmarks and still be miserable to run if every agent step takes too long.

What does the agent trace actually show?

SambaNova backed its launch with a trace from an example agent run on SambaCloud, and the numbers explain why serving speed is the pitch. The run made 41 model calls over 17 minutes and 30 seconds. It processed 5.4 million input tokens and generated just under 42,000 output tokens.

The asymmetry is the point. 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 each time. More than 91% of the input was served from cache.

So the workload is a loop: long context goes in, a short next action comes out, repeat. Per-call latency stacks up across the whole sequence, which means raw throughput determines how usable the agent feels. SambaNova frames its 200+ tokens per second as the fix for that, while the 91% cache-hit figure shows how heavily this kind of workload reuses what it already sent.

In plain terms: agents don't need a model that's fast once. They need one that's fast forty-one times in a row.

How much did post-training improve GLM-5.3?

Z.ai says it designed its post-training environments to resemble professional work, including tasks that could occupy an experienced engineer for several days. One scenario gives the model access to compute clusters, storage systems, internal documentation, codebases, and experiment results. The model must then diagnose bottlenecks across a training stack, implement optimizations, run experiments, and deliver a measurable end-to-end speedup without breaking correctness.

The reported benchmark gains span the three categories SambaNova is targeting:

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

Those are terminal work, software engineering, and general agent tasks — exactly where SambaNova aims this deployment.

Z.ai also ran comparisons on its private Code Bench. GLM-5.3 reached 31.4% at High effort, using roughly 50,000 output tokens per task. Claude Opus 4.8 scored 29.5% while consuming 120,000 tokens. That's a specific efficiency result: GLM-5.3 generated fewer than half the tokens and still scored higher. But Z.ai's own table puts Claude Fable 5 ahead overall, at 39.5% at Max effort.

The honest read: GLM-5.3 beats one Anthropic model on efficiency in one internal test, and trails another on raw score.

How do developers plug it in?

On SambaCloud, the model is called with the 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. Developers can authenticate with an API key or experiment in the playground first.

Coding-agent integration is deliberately simple. ZCode can use SambaCloud through OpenAI- or Anthropic-compatible provider settings: enter a base URL, an API key, and pick GLM-5.3 from the model menu. SambaNova says the same base-URL switch works with Claude Code, OpenCode, and Codex CLI. For Codex specifically, developers use the Responses API endpoint and declare the model's supported reasoning levels in their model catalog.

There's one migration gotcha. 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 fail.

In short: existing integrations need a small config change, not a rewrite.

Why the benchmarks and the trace tell different stories

For teams evaluating hosted agent inference, the distinction between the two datasets matters. Benchmark scores describe what a model can complete. The trace exposes the repeated serving work required to complete it.

That gap is where SambaNova positions itself. GLM-5.3's reported gains — from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1 — say the model handles the tasks. But a 41-call run that resends ever-growing context makes per-call speed and cache reuse as relevant to the user experience as any task score. A model that reasons brilliantly but takes a minute per step turns a 17-minute agent run into an afternoon.

SambaNova's deployment targets that second half of the problem. The 200+ tokens per second figure and the 91% cache hit rate are serving claims, not intelligence claims — and for agentic workloads, they may be the ones that decide whether developers stay.

The open question is whether SambaNova's speed advantage holds under real production loads beyond a supplied example trace, and whether Z.ai's efficiency edge over Opus 4.8 translates outside its own private benchmark.