Model reference

One page per model, for the models served at https://developer.amd.com.cn/radeon/api/v1.

GET /v1/models is the source of truth for what is available — a page existing here does not guarantee the model is currently published. All of them answer on /v1/chat/completions except MinerU2.5-Pro, which is document OCR and has its own endpoint.

Last updated 2026-09-23

Behaviour can change when a model moves to a different inference engine.

MiniCPM5-1B is no longer published. Its page has been removed; MiniCPM5-2B serves the same size class.

Specifications

Model	Context	Image input	Thinks with reasoning_effort omitted
DeepSeek-V4-Flash	1,048,576	❌	❌
DeepSeek-V4-Flash-Vision-Exp	1,048,576	✅	❌
DeepSeek-V4.1-Flash	1,048,576	✅	❌
Qwen3.8-Flash-Next	262,144	✅	✅
Qwen3.8-27B	131,072	✅	✅
GLM-5.3-Flash	262,144	❌	✅
MiMo-V2.6-Flash	1,048,576	✅	✅
MiniCPM5-2B	131,072	❌	❌

MiMo-V2.6-Flash is the only model here that also accepts audio input, and the only one that accepts a strict json_schema.

MinerU2.5-Pro is not in this table: it takes a PDF or an image on POST /v1/ocr, returns Markdown, and is billed per page. The parameters below do not apply to it.

Supported reasoning_effort values

The accepted tiers differ per model:

Model	Accepted values
DeepSeek-V4-Flash	none minimal low medium high xhigh max
DeepSeek-V4-Flash-Vision-Exp	none minimal low medium high xhigh max
DeepSeek-V4.1-Flash	none minimal low medium high xhigh max
Qwen3.8-Flash-Next	none low medium xhigh
Qwen3.8-27B	low medium xhigh
GLM-5.3-Flash	low medium high
MiMo-V2.6-Flash	none and the usual tiers — send none to disable thinking

The parameter may also be omitted entirely. MiniCPM5-2B does not return separated thinking, so reasoning_effort does not apply to it.

high is a 400 on Qwen3.8-27B

high is what most OpenAI-compatible clients send for “think hard”, and it is the one value that is accepted by the DeepSeek models and rejected by Qwen3.8-27B:

Unexpected reasoning effort high. Supported types are xhigh (default), medium, and low.

That model’s top tier is xhigh. minimal and max are rejected by it too.

xhigh is a 422 on GLM-5.3-Flash

GLM-5.3-Flash is the mirror image of the Qwen models: it takes the OpenAI trio and nothing else.

reasoning_effort: unknown variant `xhigh`, expected one of `low`, `medium`, `high`

So the two names for “think hard” are mutually exclusive across this catalogue: xhigh fails on GLM-5.3-Flash, high fails on Qwen3.8-27B. No single literal reaches the top tier of both. Note the status codes differ too — GLM-5.3-Flash returns 422 because the value is rejected during deserialisation, while Qwen3.8-27B returns 400 from a validator.

For one client across the thinking models, use low or medium

Those two are the intersection of every model in the table. The name of the top tier is not portable — the Qwen models spell it xhigh, GLM-5.3-Flash spells it high, the DeepSeek models accept xhigh and max and high, and they are not interchangeable. To pick per model, read GET /v1/models for what is currently published and match it against the table above.

Thinking behaviour

Model	With reasoning_effort omitted	usage.reasoning_tokens (top level)
DeepSeek-V4-Flash	does not think	✅
DeepSeek-V4-Flash-Vision-Exp	does not think	✅
DeepSeek-V4.1-Flash	does not think	✅
Qwen3.8-Flash-Next	thinks anyway	✅
Qwen3.8-27B	thinks anyway	❌
GLM-5.3-Flash	thinks anyway	❌
MiniCPM5-2B	does not think	❌

Thinking text arrives in choices[0].message.reasoning. MiniCPM5-2B answers directly, so that field is empty and reasoning_tokens is 0 — read content for its answer.

For the token count use usage.completion_tokens_details.reasoning_tokens, which every model here reports except Qwen3.8-27B — that one omits the completion_tokens_details object entirely, so its thinking tokens cannot be separated from the rest of its output. They are still counted in usage.completion_tokens, and still billed.

Thinking is on by default on both Qwen models and on GLM-5.3-Flash, and it spends max_tokens

On Qwen3.8-Flash-Next and Qwen3.8-27B a request with no reasoning_effort still thinks, and the thinking pass draws on the same max_tokens budget as the answer. A budget that was generous for a non-thinking model can come back with an empty content. Either raise max_tokens or send reasoning_effort: "low".

Writing messages

Model	system messages	Role name
DeepSeek-V4-Flash	any position, more than one allowed	system or developer
DeepSeek-V4-Flash-Vision-Exp	any position, more than one allowed	system or developer
DeepSeek-V4.1-Flash	any position, more than one allowed	system or developer
Qwen3.8-Flash-Next	exactly one, and it must come first	system
Qwen3.8-27B	at most one, and it must come first	system
GLM-5.3-Flash	any position, more than one allowed	system
MiniCPM5-2B	any position, more than one allowed	system

Only the three DeepSeek models accept developer as an alias for system. The other four reject it outright — Qwen3.8-27B with Unexpected message role., and Qwen3.8-Flash-Next, GLM-5.3-Flash and MiniCPM5-2B with a deserialisation error naming the offending index:

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

Newer OpenAI SDKs emit developer in place of system. If yours does, override it back.

The constraints on the two Qwen models come from the Jinja chat template shipped with their weights, not from a gateway rule. Qwen3.8-27B reports a violation as System message must be at the beginning.

The shape every model accepts: a single system message at index 0, with the role spelled system.

Structured output and tools

Every chat model here supports:

• response_format: {"type": "json_object"}
• function calling via tools + tool_choice

parallel_tool_calls is accepted everywhere, but emitting several tool_calls in one turn is the model’s own choice — do not build on it.

Image input

Four models take images; put an image_url content part in the content array:

Model	Per-image metering
DeepSeek-V4-Flash-Vision-Exp	usage.prompt_tokens_details.image_tokens, 115 in the sample
DeepSeek-V4.1-Flash	usage.prompt_tokens_details.image_tokens, 202 in the sample
Qwen3.8-Flash-Next	usage.prompt_tokens_details.image_tokens, 144 in the sample
Qwen3.8-27B	usage.prompt_tokens_details.multimodal_tokens.image, 72 in the sample

Qwen3.8-27B nests the image count one level deeper

Three of the four report prompt_tokens_details.image_tokens. Qwen3.8-27B reports prompt_tokens_details.multimodal_tokens.image instead, so code that tallies image usage across models has to read both shapes.

{
  "model": "DeepSeek-V4-Flash-Vision-Exp",
  "messages": [{
    "role": "user",
    "content": [
      { "type": "image_url", "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } },
      { "type": "text", "text": "What does this image say?" }
    ]
  }]
}

Check the table above before sending images; the remaining chat models are text-only. To pull text out of a PDF or a scan, use MinerU2.5-Pro — it returns Markdown and is billed per page.

Writing one client for every model

The shape every chat model here accepts:

• a single system message at index 0, with the role spelled system
• reasoning_effort of low or medium, or omitted entirely — not high (400 on Qwen3.8-27B) and not xhigh (422 on GLM-5.3-Flash)
• read thinking text from choices[0].message.reasoning
• read thinking tokens from usage.completion_tokens_details.reasoning_tokens (Qwen3.8-27B omits it)
• use response_format: {"type": "json_object"} for JSON
• keep content textual unless the target model supports images
• size max_tokens against the target model’s window — they span 131,072 to 1,048,576, an eightfold difference

True for every model

Accepted parameters	temperature, max_tokens, top_p, stream, response_format, tools, tool_choice, parallel_tool_calls, reasoning_effort
Silently dropped	stop, seed, logit_bias, logprobs, top_logprobs, top_k, min_p, repetition_penalty
How to turn thinking on	reasoning_effort (or the equivalent reasoning.effort)
Reported tokenizer	GPT
Stability	experimental

Parameters outside the accepted list are stripped before the request reaches the inference backend — no error, no effect. If you need them, use a dedicated endpoint, which forwards the request body to vLLM or SGLang untouched.