Model catalog

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 five, which have their own endpoints: MinerU2.5-Pro (document OCR), OmniVoice (text-to-speech), and Clef, d3 and d3-flash (decisions).

Last updated 2026-10-19

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

New: Gemma-4-31B-IT. The tables below now also cover MiniCPM-V-4.7-1B and MiniCPM-V-4.7-35B-A3B.

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

Model ID and card title

Send the model ID in model, exactly as written. Copy it from the Model field on the model’s card in Token Factory, or from id in GET /v1/models. The card title and the name field can differ from the ID, and neither is accepted in its place:

Model ID (send this)	Card title	name in /v1/models
DeepSeek-V4-Flash	DeepSeek-V4-Flash-0731	DeepSeek-V4-Flash
DeepSeek-V4.1-Flash	DeepSeek-V4.1-Flash-1017	DeepSeek-V4.1-Flash
Gemma-4-31B-IT	Gemma-4-31B-IT	Gemma 4 31B

For every other model, the ID, the card title and name are the same.

Specifications

Thinks by default means the model thinks even when the request doesn’t ask it to. See Thinking.

Model	Context	Image input	Thinks by default
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	262,144	✅	✅
GLM-5.3-Flash	262,144	❌	✅
MiMo-V2.6-Flash	1,048,576	✅	✅
MiniCPM5-2B	131,072	❌	❌
MiniCPM-V-4.7-1B	262,144	✅	❌
MiniCPM-V-4.7-35B-A3B	262,144	✅	❌
Gemma-4-31B-IT	262,144	✅	❌

Five models are not in this table, and the parameters below do not apply to them:

Model	Endpoint	Takes	Returns	Billed per
MinerU2.5-Pro	POST /v1/ocr	a PDF or an image	Markdown	page
OmniVoice	POST /v1/audio/speech	text, up to 4,096 characters	an audio file	input character
Clef	POST /v1/systemone	a state, optional images, typed questions	a probability for every answer	input token
d3	POST /v1/systemone	a state, optional images and videos, typed questions	a probability for every answer	input token
d3-flash	POST /v1/systemone	a state, optional images and videos, typed questions	a probability for every answer	input token

Usage rates

Token Factory meters usage in points (pts). Points aren’t money and are separate from your credits; see Usage and quota. The rates below were read from the model cards on 2026-10-19 and are in pts per million tokens unless noted. Rates can change, so check the model’s card for the current values.

Model	Input	Output	Cache read
DeepSeek-V4-Flash	0.14	0.28	0.0028
DeepSeek-V4-Flash-Vision-Exp	0.14	0.28	0.0028
DeepSeek-V4.1-Flash	0.28	0.56	0.0056
Qwen3.8-Flash-Next	0.14	0.28	0.0028
Qwen3.8-27B	0.035	0.07	0.0007
GLM-5.3-Flash	0.28	0.56	0.0056
MiMo-V2.6-Flash	0.14	0.28	0.0028
MiniCPM5-2B	0.035	0.07	0.0007
MiniCPM-V-4.7-1B	0.035	0.07	0.0007
MiniCPM-V-4.7-35B-A3B	0.035	0.07	0.0007
Gemma-4-31B-IT	0.035	0.07	0.0007
Clef	0.035	—	—
d3	0.035	—	—
d3-flash	0.035	—	—
OmniVoice	0.035 per million characters	—	—
MinerU2.5-Pro	0.001 per page	—	—

Three things on the cards differ from what is billed:

• The Clef, d3, d3-flash and OmniVoice cards also show a Cache read rate. Those models are not billed for cache reads.
• The OmniVoice card labels its input rate per 1M tokens. It is billed per million input characters.
• The MinerU2.5-Pro card shows no rate. It is billed per page processed.

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 minimal low medium high xhigh max

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

MiMo-V2.6-Flash accepts every value but doesn’t act on it: in tests on 2026-10-19, every value, none included, returned the same short reasoning. You can’t turn its thinking off.

MiniCPM-V-4.7-1B, MiniCPM-V-4.7-35B-A3B and Gemma-4-31B-IT turn thinking on with chat_template_kwargs: {"enable_thinking": true} instead. How they treat reasoning_effort has not been verified.

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

Model	Without a thinking parameter	Top-level reasoning_tokens
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	❌
MiMo-V2.6-Flash	thinks anyway	❌
MiniCPM5-2B	does not think	❌
MiniCPM-V-4.7-1B	does not think	not verified
MiniCPM-V-4.7-35B-A3B	does not think	not verified
Gemma-4-31B-IT	does not think	not verified

The Qwen3.8-27B usage values were measured on its earlier vLLM deployment.

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. There are two exceptions. Qwen3.8-27B omitted the completion_tokens_details object on its earlier deployment, and MiMo-V2.6-Flash reports 0 there even when it thinks. Their thinking tokens cannot be separated from the rest of the 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
MiMo-V2.6-Flash	any position, more than one allowed	system or developer
MiniCPM5-2B	any position, more than one allowed	system
MiniCPM-V-4.7-1B	one, first; other placements not verified	system
MiniCPM-V-4.7-35B-A3B	one, first; other placements not verified	system
Gemma-4-31B-IT	one, first; other placements not verified	system

The three DeepSeek models and MiMo-V2.6-Flash accept developer as an alias for system. Qwen3.8-27B rejects it with Unexpected message role., and Qwen3.8-Flash-Next, GLM-5.3-Flash and MiniCPM5-2B reject it with a deserialisation error naming the offending index:

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

The developer role has not been verified on the MiniCPM-V-4.7 models or Gemma-4-31B-IT. 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

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

Model	Where image usage is reported
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, on its earlier deployment
MiMo-V2.6-Flash	usage.prompt_tokens_details.image_tokens
MiniCPM-V-4.7-1B	usage.prompt_tokens_details.image_tokens
MiniCPM-V-4.7-35B-A3B	usage.prompt_tokens_details.image_tokens
Gemma-4-31B-IT	not verified

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

On its earlier deployment, Qwen3.8-27B reported prompt_tokens_details.multimodal_tokens.image instead of prompt_tokens_details.image_tokens. Code that tallies image usage across models should 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.

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 omitted it on its earlier deployment, and MiMo-V2.6-Flash reports 0)
• 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

Common to every model

Accepted parameters	temperature, max_tokens, top_p, stream, response_format, tools, tool_choice, parallel_tool_calls, reasoning_effort
Forwarded unchanged	chat_template_kwargs
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); chat_template_kwargs.enable_thinking on the MiniCPM-V-4.7 models and Gemma-4-31B-IT
Reported tokenizer	GPT
Stability	experimental

Other parameters are stripped before the request reaches the inference backend — no error, no effect. If you need them, serve the model yourself as a model service, which forwards the request body to vLLM untouched.