MiniCPM-V-4.7-35B-A3B

POST /v1/chat/completions model: MiniCPM-V-4.7-35B-A3B

About the model

The weights served here are OpenBMB/MiniCPM-V-4.7-35B-A3B, the V4.7 family’s 35B-total, approximately 3B-active MoE vision-language model, with a Qwen3.5-MoE text backbone. It accepts text and images and answers without thinking unless enabled explicitly. The smaller dense option is MiniCPM-V-4.7-1B. Neither model is the separate text-only MiniCPM5-2B; A3B does not mean the whole model is 3B.

Architecture

From the official weights’ pinned config.json. Layer and attention dimensions below refer to text_config:

Architecture	MiniCPMV4_7ForConditionalGeneration, model_type = minicpmv4_7
Text backbone	Qwen3.5-MoE, model_type = qwen3_5_moe_text
Layers	40 — 30 linear-attention + 10 full-attention; every fourth layer is full attention
Hidden dimension	2,048
Full attention	16 Q heads, 2 KV heads; head dimension 256
Linear attention	16 key heads, 32 value heads; key/value head dimensions 128; convolution kernel 4
Mixture of Experts	256 experts, 8 selected per token
Intermediate dimension	Routed experts: 512; shared expert: 512
Vocabulary	248,144
Native context	262,144
Precision	bfloat16 — unquantised weights

The separate vision_config describes the vision tower:

Model type	minicpmv4_7_vision
Layers / hidden dimension / heads	27 / 1,152 / 16
Intermediate dimension	4,304
Patch size	14
Configured image size	980 — not an uploaded-image size limit

The outer configuration sets the default visual compression to downsample_mode = 16x.

On Radeon Cloud

Token Factory, China site — verified 2026-10-13

These service behaviours apply to https://developer.amd.com.cn/radeon/api/v1, not every regional deployment or model service. They were verified on 2026-10-13, when the model ran on SGLang. It has been served by vLLM since 2026-10-16. Current availability comes from GET /v1/models, and current usage rates are on the model’s card in the console.

At a glance

Context length	262,144 tokens
Input modalities	Text, image
Streaming	✅
Tool calling	✅ — including tool-result continuation and streamed calls
JSON output	✅ json_object; strict json_schema unavailable for this model here
Thinking	✅ — off by default, enable explicitly
Inference engine	vLLM
Stability	experimental

messages

Use ordinary user / assistant turns. If needed, put one system message first. Multiple or later system messages and the developer role have not been qualified for this model. Images go in a user message’s content array, alongside text parts.

Thinking

Thinking is off by default. To enable it, send this top-level request field:

{"chat_template_kwargs": {"enable_thinking": true}}

With the Python OpenAI SDK, use extra_body={"chat_template_kwargs": {"enable_thinking": True}}. Set enable_thinking to false to disable it explicitly. A reasoning_effort tier or separate thinking-token-budget contract has not been established for this model.

Thinking text, non-streaming	choices[].message.reasoning
Thinking text, SSE	choices[].delta.reasoning
Final answer	choices[].message.content, or choices[].delta.content in SSE
Separate thinking-token count, when supplied	usage.completion_tokens_details.reasoning_tokens or usage.reasoning_tokens

Read reasoning, not reasoning_content, and keep it separate from the final answer. Absent/null text fields are valid. For SSE, enable stream: true; request final usage with stream_options: {"include_usage": true} and handle usage-only chunks with choices: []. Keep usage counts as returned: do not sum the two thinking counters or add them again to totals. A missing counter means unavailable, not necessarily zero. Thinking SSE was tested; non-streaming field names follow the gateway contract but were not separately re-tested in this round.

Image input

Single-image and two-image requests passed functional tests. Use an image_url content part:

{
  "role": "user",
  "content": [
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64_PNG>"}},
    {"type": "text", "text": "Describe the image and transcribe any visible text."}
  ]
}

Replace <BASE64_PNG> with the image’s base64 data and match the MIME type to its format. For multiple images, repeat the image_url part in question order. This is not an unlimited-image or OCR-accuracy guarantee. Image usage is reported in usage.prompt_tokens_details.image_tokens; visual tokens consume context, and base64 increases request size.

The service uses 16x visual compression. mm_processor_kwargs, including downsample_mode="4x", is not forwarded; sending it does not change the compression profile.

Tools

tools with tool_choice="auto", "required", or a named function passed tests, as did tool-result continuation and streamed tool calls. The model requests a call; your application executes it. Validate names and arguments. Append the assistant message containing tool_calls to the conversation, then append each result as a role: "tool" message with its matching tool_call_id before requesting the final answer.

Only execute complete calls after finish_reason="tool_calls". For SSE, accumulate delta.tool_calls by choice and tool-call index before parsing argument JSON. Check the final answer’s finish_reason as well: length is not a complete answer. parallel_tool_calls is not forwarded; false cannot guarantee one call. Handle all returned calls and execute them serially in your application when needed.

Structured output

response_format	Result for this model on this deployment
{"type":"json_object"}	✅ — ask for JSON in the prompt and parse choices[0].message.content
{"type":"json_schema","json_schema":{...}}	❌ — returns 400 for MiniCPM-V-4.7-35B-A3B here

JSON object mode and JSON tool arguments do not enforce a response schema. Check for a complete response before parsing, then validate required fields and types in your application. This restriction describes this MiniCPM deployment, not every model in Token Factory.

Limits

Context window	262,144 tokens, prompt plus generated output
Configured output limit	32,768 tokens, also constrained by remaining context
Input	Text and images; video_url returns 400 here; audio unverified
JSON output	json_object; no strict json_schema for this model here
Thinking switch	chat_template_kwargs.enable_thinking
Request overrides	parallel_tool_calls and mm_processor_kwargs not forwarded

Configured context is not full-window qualification

System text, history, tool definitions, the chat template and images consume context; thinking also consumes the output budget. Do not send 256K input tokens and expect additional output. Full 256K public end-to-end requests have not been qualified. Long agent loops, thinking combined with structured output, and this model’s Anthropic/Responses compatibility are also unverified.

For normal text/image answers and the final answer after tool execution, require finish_reason="stop" before treating output as complete. In SSE, inspect the terminal choice; length or a missing finish reason must not silently pass as a finished answer. Large requests can encounter ingress, timeout and rate-limit constraints.

Example

Use the exact model ID without openbmb/. Set RADEON_API_KEY to your Token Factory API key in the environment; never commit a real key.

curl --fail-with-body https://developer.amd.com.cn/radeon/api/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniCPM-V-4.7-35B-A3B",
    "messages": [{"role": "user", "content": "Calculate 17 + 25. Reply with the number only."}],
    "temperature": 0,
    "max_tokens": 64
  }'

Read choices[0].message.content after checking finish_reason; token counts are in usage. Usage is metered in points (pts) per million tokens. The current rates are in the usage rates table and on the model’s card. Points aren’t money and are separate from your credits.