GLM-5.3-Flash is live on CloudCode.ONE — and it can finally see

Every GLM flagship I've added to this gateway so far has had the same asterisk attached: great at code, blind as a bat. That's why CloudCode.ONE has automatic vision routing — you send an image to a text-only model, we quietly hand that part to something that can actually look at it, and your agent loop doesn't break.

GLM-5.3-Flash removes the asterisk. It's the first natively multimodal model in the GLM-5 series, and vision isn't bolted on as a separate captioning step — it's wired into the coding loop itself. The model can render its own frontend, look at the screenshot, notice the padding is wrong, and go fix it.

It's available on CloudCode.ONE now as glm-5.3-flash.

The short version

320B total parameters, 18B activated. Roughly half the active parameters of the GLM-4.5 generation, and 45 layers instead of 92.
1M-token context, with a hybrid linear + sparse attention architecture built specifically to make long context cheap to serve.
Native image input — multiple images per request, URL or base64.
Thinking is always on. There's no disable switch on this one (more on that below, it matters for your client config).
Function calling, structured output, streaming, context caching all supported.

The benchmarks that actually matter

Z.ai reports GLM-5.3-Flash scoring 57 on the Artificial Analysis Intelligence Index v4.1.1 at roughly $0.045 per task at discounted rates — a score that until recently lived about an order of magnitude higher up the price curve. That's the whole pitch in one line: the intelligence tier moved, the price tier didn't.

Against GLM-5.2, which many of you are running today, the coding and agentic deltas are not subtle:

Benchmark	GLM-5.2	GLM-5.3-Flash
DeepSWE v1.1	46.2	63.4
AutomationBench	26.2	48.8

On Z.ai's in-house coding evaluation (Z.ai Code Bench v1.0, run through Claude Code), GLM-5.3-Flash beats GLM-5.2 at every effort level, and at maximum effort lands at 29.0 against Claude Opus 4.8's 29.5. Take vendor-run in-house benchmarks with the usual grain of salt, but the direction is consistent across six external coding and agentic benchmarks too.

Worth noting for the skeptics: this model was road-tested anonymously as ox-alpha on OpenCode and OpenRouter before release. If you tried an unnamed model last week and wondered why it was punching above its weight, that was probably this.

Why it's cheap (the part I find genuinely interesting)

Most "flash" models get cheap by being small. This one gets cheap by restructuring attention.

GLM-5.3-Flash combines linear attention — which models local dependencies through a running state rather than an ever-growing cache — with sparse attention that pulls in global context through a lightweight index. Compared to GLM-5.3, that cuts attention compute by about 3× and KV cache size by about 4.4×. At a 1M-token context window, KV cache is the thing that kills your serving economics, so this is the difference between "long context exists" and "long context is affordable."

The serving side is a story too: Z.ai is running this on a large cluster of domestically developed Chinese AI chips, with a custom SGLang-based inference engine, and reports reaching per-token cost comparable to mainstream NVIDIA hardware. For those of us buying inference rather than building it, that's a supply-side diversification worth paying attention to.

What changes for CloudCode.ONE users

If you're using vision routing today: nothing breaks. Automatic vision routing stays exactly where it is — it's still doing real work for DeepSeek V4 and the text-only GLM models, and it stays the default safety net for anything that can't take an image. But if you're on glm-5.3-flash, your images now go straight to the model that's doing the reasoning, in the same request, in the same turn. No handoff, no summarized description of your screenshot standing in for the screenshot.

That distinction matters more than it sounds. Routed vision gives your model a description of what it built. Native vision lets the model form its own judgment about whether the spacing is off — and then iterate against it. For frontend work, Godot prototypes, Blender scenes, or anything where the deliverable is something a human looks at rather than reads, that loop is the whole ballgame.

If you're on GLM-5.2: switching is a one-line model-string change. Same Anthropic-compatible endpoint, same auth, same tooling.

Using it

Anthropic-compatible endpoint, text only:

curl https://api.cloudcode.one/v1/messages \
  -H "x-api-key: $CLOUDCODE_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "glm-5.3-flash",
    "max_tokens": 4096,
    "messages": [
      { "role": "user", "content": "Refactor this module to remove the circular import." }
    ]
  }'

With an image in the same turn:

{
  "model": "glm-5.3-flash",
  "max_tokens": 4096,
  "messages": [{
    "role": "user",
    "content": [
      { "type": "image", "source": { "type": "url", "url": "https://example.com/design-reference.png" } },
      { "type": "text", "text": "Build this as a Next.js page. Match the spacing and type scale." }
    ]
  }]
}

Add more image blocks for multi-page references — screenshot sets are exactly the use case this model was trained for.

Claude Code

export ANTHROPIC_BASE_URL="https://api.cloudcode.one"
export ANTHROPIC_AUTH_TOKEN="your-cloudcode-key"
export ANTHROPIC_MODEL="glm-5.3-flash"
claude

Harness

Pick glm-5.3-flash from the model selector. Vision input works out of the box on Windows and Ubuntu builds.

Settings worth knowing

Z.ai's recommended parameters for this model differ from what you may have tuned for GLM-5.2:

temperature: 1, top_p: 0.95
reasoning_effort: max for agentic work
Thinking cannot be disabled. Only the enabled mode is supported. If you have a client config that explicitly turns thinking off, it needs updating.
clear_thinking: false is recommended — which lines up neatly with the preserved-thinking handling the gateway already does for GLM models across long agentic sessions.
For streaming, enable both stream and tool_stream.

It's not only a coding model

The vision integration pushes this into work that isn't code at all. Z.ai's own recommended workflows include generating and then visually inspecting PPTX, DOCX, XLSX and PDF deliverables — catching text overflow, misaligned elements and cropped images by actually rendering the page and looking at it. Also: Blender scenes iterated through fixed-camera renders, Godot game prototypes tested for a complete gameplay loop, and computer-use flows where the model operates an interface with no structured API available.

If you've been running a separate model for document generation, it's worth a bake-off.

All posts · CloudCode.ONE home