GLM-5.3-Flash

POST /v1/chat/completions model: GLM-5.3-Flash

About the model

The weights served here are GLM-5.3-Flash, Z.AI’s speed-oriented member of the GLM-5.3 line — a sparse Mixture-of-Experts model that thinks before answering, and does so without being asked.

Architecture

From the shipped config.json:

Architecture	Glm5NextForConditionalGeneration, model_type = glm5_next
Layers	45 — the first 3 dense, the rest MoE
Hidden dimension	4,096
Attention	64 query heads, 64 KV heads — full MHA, not grouped
Dense intermediate	12,288
Mixture of Experts	288 routed experts + 1 shared, 8 routed activated per token, expert intermediate 2,048
Multi-token prediction	1 layer
Vocabulary	154,880
Native context	1,048,576
Quantisation	FP8, e4m3, block [128, 128], dynamic activation scaling
Licence	MIT

Served at 262,144, not the native 1,048,576

The weights declare a 1M-token window, but this endpoint runs them with --max-model-len 262144. 262,144 is what you get here; requests are measured against that, not against the number in config.json.

The weights carry a vision encoder — this endpoint does not use it

config.json includes a glm5_next_vision tower (24 layers, 448 px, patch 14) and image/video token ids, so the release is multimodal. This deployment serves text only. See Image input for what an image request actually returns.

On this endpoint

At a glance

Context length	262,144 tokens
Input modalities	text only
Streaming	✅
Tool calling	✅
JSON output	✅ json_object
Thinking	✅ — on unless you turn it down
Inference engine	vLLM
Stability	experimental

messages

A system message may sit at any position, and there may be more than one — unlike Qwen3.8-Flash-Next, which allows exactly one and insists it comes first.

Spell the role system, not developer

The accepted roles are system, user, assistant, tool. Newer OpenAI SDKs emit developer in place of system; if yours does, override it back to system. There is no fallback — the request is rejected while the body is being deserialised, with a 422:

Failed to deserialize the JSON body into the target type: messages[0]: unknown role: developer

Thinking

Omitting reasoning_effort does not disable thinking. A plain request already returns a populated reasoning. How long the model thinks tracks the prompt rather than the tier — a one-line question may produce a few dozen reasoning tokens, a puzzle several thousand.

Accepted values:

Tier	Notes
omitted	still thinks
low
medium
high	the longest tier

xhigh is a 422 here — this model is the exception

Every other thinking model on this platform takes xhigh as its top tier, and Qwen3.8-Flash-Next and Qwen3.8-27B reject high. GLM-5.3-Flash is the mirror image: high works, xhigh does not.

Failed to deserialize the JSON body into the target type: reasoning_effort: unknown variant `xhigh`, expected one of `low`, `medium`, `high` at line 1 column 131

The status is 422, not the 400 that Qwen3.8-27B returns for the same class of mistake — the value is rejected while the request body is being deserialised, before any validator sees it.

none, minimal and max produce the same 422 with their own name in the message. Only low, medium and high are accepted — the OpenAI trio, and nothing else.

Portable code that has to run against both this model and the Qwen models cannot share one literal for “think hard”: send high here and xhigh there, or omit the parameter on both.

There is no reasoning_effort value that turns thinking off — none is not in the enum. To keep a response short, cap max_tokens instead, and read the answer from content rather than reasoning.

Where the output lands:

Thinking text	choices[0].message.reasoning
Token count	usage.completion_tokens_details.reasoning_tokens

Thinking consumes your max_tokens budget

Reasoning tokens are billed and counted as output. A small max_tokens can be spent entirely on thinking, leaving content as null and finish_reason as length. Budget for both, or lower the tier.

Image input

Not available on this endpoint, despite the vision tower in the weights. An image_url part comes back as a 400 whose message describes a server-side path setting rather than the real reason:

Invalid `--allowed-local-media-path`: The path <path> does not exist.

Send text only. For images use DeepSeek-V4-Flash-Vision-Exp, DeepSeek-V4.1-Flash or either Qwen3.8 model.

Limits

Context window	262,144 tokens, counted as a total budget — prompt plus output, not an output-only allowance
Input	text only; for images use DeepSeek-V4.1-Flash
JSON output	response_format: {"type": "json_object"}
Thinking tiers	reasoning_effort (or the equivalent reasoning.effort) — low, medium, high

max_tokens is capped against the total budget, prompt included:

This model's maximum context length is 262144 tokens. However, you requested 128000 output tokens and your prompt contains at least 134145 input tokens, for a total of at least 262145 tokens. Please reduce the length of the input prompt or the number of requested output tokens.

That one is a 400, unlike the 422 above.

Example

curl https://developer.amd.com.cn/radeon/api/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "GLM-5.3-Flash",
    "reasoning_effort": "high",
    "messages": [
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  }'

Swap high for xhigh and the same request becomes a 422 on this model — see Thinking.