DeepSeek-V4-Flash-Vision-Exp
DeepSeek-V4-Flash-Vision-Exp
About the model
Section titled “About the model”A vision tower bolted onto the DeepSeek-V4-Flash
weights; the Exp in the name is experimental. It has a one-million-token context and takes
images, like DeepSeek-V4.1-Flash and
MiMo-V2.6-Flash.
The language side shares its architecture with the text model, field for field. The table below is
read from the config.json of the weights this endpoint actually loads.
| Architecture | DeepseekV4ForCausalLM, model_type = deepseek_v4 |
| Layers | 43 |
| Hidden size | 4,096 |
| Attention | 64 query heads, 1 KV head, head dim 512 (MLA) |
| Mixture of experts | 256 routed + 1 shared, 6 activated per token, expert intermediate 2,048 |
| Sparse index | a DSA indexer with index_n_heads / index_topk |
| Sliding window | 128 |
| Vocabulary | 129,280 |
| Native context | 1,048,576 |
| Quantisation | FP8 e4m3, block size [128, 128], dynamic activation scaling, ue8m0 scale format |
The vision tower (same config.json, expressed as flat vision_* keys rather than a nested
vision_config):
| Layers / width / heads | 32 layers, 1,024 wide, 16 heads |
| Patch size | 14 |
| Intermediate size | 2,816 |
| Downsample ratio | 3 |
| Per-image token ceiling | 384 |
| Minimum pixels | 147,456 |
| Maximum width:height ratio | 8 |
vision_max_n_token = 384 is a hard ceiling: a larger image is compressed to 384 tokens, so very
high-resolution inputs do not cost linearly more — and do not carry more detail either.
On Radeon Cloud
Section titled “On Radeon Cloud”At a glance
Section titled “At a glance”| Context length | 1,048,576 tokens |
| Input modalities | text + images |
| Streaming | ✅ |
| Tool calling | ✅ |
| JSON output | ✅ json_object |
| Thinking | ✅ — off by default, ask for it explicitly |
| Stability | experimental |
Image input
Section titled “Image input”| How to send | an {"type":"image_url","image_url":{"url":"data:image/png;base64,…"}} part in content |
| Metering | usage.prompt_tokens_details.image_tokens |
| Per-image ceiling | 384 tokens (above) |
curl https://developer.amd.com.cn/radeon/api/v1/chat/completions \ -H "Authorization: Bearer $RADEON_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "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?" } ] }] }'messages
Section titled “messages”As permissive as the text model: a system message may sit at any position, there may be more than
one, and the role may be spelled either system or developer.
This differs from Qwen3.8-27B, which accepts
at most one system, and only at the start. For one client driving both, follow the stricter shape.
Thinking
Section titled “Thinking”Off by default. With reasoning_effort omitted, reasoning_tokens comes back as 0; you have to
ask for thinking explicitly.
All six reasoning_effort values are accepted (minimal, low, medium, high, xhigh, max,
plus none and omission) — the most permissive model on this endpoint.
| Thinking text | choices[0].message.reasoning |
| Token count | usage.completion_tokens_details.reasoning_tokens |
| Top-level alias | usage.reasoning_tokens — this model does emit it |
Limits
Section titled “Limits”| Context window | 1,048,576 tokens, counted as prompt plus output |
| JSON output | response_format: {"type": "json_object"} |
| Turning thinking on | reasoning_effort (or the equivalent reasoning.effort) |