Qwen3.5 Omni Flash

alibaba/qwen3.5-omni-flash — a llm model by Alibaba. Call it through the gateway’s standard OpenAI-compatible door using the catalog slug; the full parameter schema is reproduced in full below.

Catalog slug	alibaba/qwen3.5-omni-flash
Kind	llm
Vendor	Alibaba
Context window	1,000,000 tokens
Native model id	qwen3.5-omni-flash

How to call

Through the Sociaro SDK / OpenAI-compatible door

from sociaro_ai import Sociaro

client = Sociaro(api_key="gw_live_...")   # base_url defaults to https://api.sociaro.com

resp = client.chat.completions.create(
    model="alibaba/qwen3.5-omni-flash",          # always the catalog slug
    messages=[{"role": "user", "content": "Hello"}],
    # any parameter from the schema below is passed through unchanged
)
print(resp.choices[0].message.content)

TypeScript: client.chat.completions.create({ model: "alibaba/qwen3.5-omni-flash", messages: [...] }) — see the TypeScript SDK.

Through Alibaba DashScope’s native API (advanced)

Optional: send Alibaba DashScope’s native request to the /alibaba prefix, using the native model id qwen3.5-omni-flash. The gateway injects the upstream Alibaba DashScope credential and forwards the body unchanged (catalog, pricing and entitlements still apply).

curl https://api.sociaro.com/alibaba/<native-path> \
  -H "Authorization: Bearer $SOCIARO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "qwen3.5-omni-flash", ... }'

The exact path, request body and response shape are documented in full in the schema below.

Parameters — full schema

The complete, one-to-one API schema for this model, transcribed from the official source. Nothing omitted. Examples in the schema may use a representative model of the family — for this model use the slug alibaba/qwen3.5-omni-flash (or the native id qwen3.5-omni-flash for the native API) as shown in How to call above.

Coverage map (per-model)

model slug	hosting	input modalities	output modalities	thinking	web search	voice cloning	default voice	voice count
qwen3.5-omni-plus	Model Studio (OpenAI-compatible)	text + (image / audio / video), combined	text, audio	no	yes (enable_search, search_strategy=agent)	yes (not snapshots)	Tina	55
qwen3.5-omni-flash	Model Studio (OpenAI-compatible)	text + (image / audio / video), combined	text, audio	no	yes (enable_search, search_strategy=agent)	yes (not snapshots)	Tina	55
qwen3.5-omni-plus-2026-03-15 (snapshot)	Model Studio	as above	text, audio	no	yes	no	Tina	55
qwen3.5-omni-flash-2026-03-15 (snapshot)	Model Studio	as above	text, audio	no	yes	no	Tina	55
qwen3-omni-flash (stable)	Model Studio	text + exactly one of (image / audio / video); 1 media file	text, audio (no audio in thinking)	yes (enable_thinking)	no	no	Cherry	17 (stable == 2025-09-15 set)
qwen3-omni-flash-2025-12-01 (snapshot)	Model Studio	same as flash	text, audio (no audio in thinking)	yes	no	no	Cherry	49
qwen3-omni-flash-2025-09-15 (snapshot)	Model Studio	same as flash	text, audio (no audio in thinking)	yes	no	no	Cherry	17
qwen-omni-turbo (legacy, deprecated)	Model Studio	text + one other modality	text, audio	no	no	no	Chelsie	4
qwen3-omni-30b-a3b-captioner	Open-source self-host (HF / ModelScope / vLLM); not a Model Studio hosted endpoint	audio only, single file, single-turn, no text prompt	text only (no audio, no voice)	n/a	n/a	n/a	—	—

Model names, context windows, exact pricing and the full snapshot list are confirmed in the Model Studio console. Regions: Singapore and Beijing (API keys differ per region).

---

Critical per-model differences (read first)

1. Combined multimodal input (image + audio + video, or any mix, in ONE user message) is supported only by the Qwen3.5-Omni series. qwen3-omni-flash and qwen-omni-turbo accept text + exactly one other modality and only one media file per request.
2. Thinking mode (enable_thinking) exists only for qwen3-omni-flash within the Omni family. In thinking mode, audio output is not supported — modalities must be ["text"] and the audio parameter is ignored. Reasoning is streamed in delta.reasoning_content.
3. Web search (enable_search) is supported only by the Qwen3.5-Omni series (incl. its dated snapshots). For Omni the applicable search_strategy value is agent.
4. vl_high_resolution_images and the per-frame/per-image pixel knobs that raise the cap to the 16384-token resolution apply to Qwen3.5-Omni. qwen3-omni-flash and qwen-omni-turbo do not support vl_high_resolution_images. Image token density is 1 token per 32×32 px for the Qwen3.5-Omni series and Qwen3-Omni-Flash, and 1 token per 28×28 px for Qwen-Omni-Turbo.
5. Streaming is mandatory for every hosted Qwen-Omni call: stream=true is required, otherwise the request errors.
6. Output audio is always PCM 24000 Hz, mono, 16-bit, delivered as base64 fragments in delta.audio.data to be concatenated and decoded. Output audio.format is wav only.
7. qwen3-omni-30b-a3b-captioner is open-source self-host, not a Model Studio OpenAI-compatible endpoint. Audio-in / text-out only, single audio file, no text prompt, single-turn.
8. Input limits diverge sharply by series (audio ≤3 h vs ≤20 min vs ≤3 min; video ≤1 h vs ≤150 s vs ≤40 s; files: Qwen3.5-Omni up to 512/2048 vs Flash & Turbo only 1) — see the input-modality tables below.

---

Endpoint / transport

• Mode: OpenAI-compatible only (DashScope native mode is not supported for Qwen-Omni).
• Endpoint: POST {base_url}/chat/completions
• base_url (Singapore example): https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1 (Beijing region uses a different host/key.)
• Auth: Authorization: Bearer $DASHSCOPE_API_KEY
• Min SDK: OpenAI Python ≥ 1.52.0, Node.js ≥ 4.68.0.
• Hard rule: every hosted Qwen-Omni call MUST set stream: true. Recommended: stream_options: {"include_usage": true} to receive usage in the final chunk.
• extra_body transport: the following are NOT standard OpenAI params and (with the Python/Node SDK) must be passed inside extra_body: top_k, repetition_penalty, enable_thinking, thinking, thinking_budget, preserve_thinking, reasoning_effort, tool_stream, enable_code_interpreter, enable_search, search_options, vl_high_resolution_images. (Raw HTTP/curl puts them at the top level of the JSON body.)

---

Request body shape (full, illustrative)

{
  "model": "qwen3.5-omni-plus",
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "Describe the image and the audio." },
        { "type": "image_url", "image_url": { "url": "https://.../pic.jpeg" }, "min_pixels": 24576, "max_pixels": 1310720 },
        { "type": "input_audio", "input_audio": { "data": "https://.../clip.wav", "format": "wav" } },
        { "type": "video_url", "video_url": { "url": "https://.../movie.mp4" }, "fps": 2.0 },
        { "type": "video", "video": ["https://.../frame1.jpg", "https://.../frame2.jpg"], "fps": 2.0 },
        { "type": "text", "text": "concise please", "cache_control": { "type": "ephemeral" } }
      ]
    }
  ],
  "modalities": ["text", "audio"],
  "audio": { "voice": "Tina", "format": "wav" },
  "stream": true,
  "stream_options": { "include_usage": true },
  "temperature": 1.0,
  "top_p": 1.0,
  "top_k": 20,
  "presence_penalty": 1.5,
  "repetition_penalty": 1.0,
  "max_completion_tokens": 2000,
  "n": 1,
  "seed": 1234,
  "stop": null,
  "logprobs": false,
  "top_logprobs": 0,
  "response_format": { "type": "text" },
  "tools": [],
  "tool_choice": "auto",
  "parallel_tool_calls": false,
  "enable_thinking": false,
  "thinking_budget": 2048,
  "enable_search": false,
  "search_options": { "forced_search": false, "search_strategy": "agent", "enable_search_extension": false },
  "vl_high_resolution_images": false
}

modalities accepts only ["text","audio"] or ["text"]. Thinking mode (qwen3-omni-flash) forces ["text"]. enable_thinking applies only to qwen3-omni-flash; enable_search/search_options only to Qwen3.5-Omni; vl_high_resolution_images only to Qwen3.5-Omni.

---

Request parameters (one-to-one with the OpenAI-compatible Chat Completions reference)

name	type	required	default	allowed values / range	description & Omni notes
model	string	yes	—	hosted: qwen3.5-omni-plus, qwen3.5-omni-flash (+ snapshots …-2026-03-15), qwen3-omni-flash (+ snapshots …-2025-12-01, …-2025-09-15), qwen-omni-turbo (legacy); open-source: qwen3-omni-30b-a3b-captioner	Model name. For exact names/pricing check the console.
messages	array	yes	—	objects with role ∈ {system,user,assistant,tool}	Conversation history in chronological order. See message-types below.
stream	boolean	yes for Qwen-Omni (must be true)	false	true/false; Qwen-Omni rejects non-stream	Streaming output. Non-streaming raises an error for Qwen-Omni.
stream_options	object	no	—	{ "include_usage": boolean }	Effective only when stream=true. include_usage:true ⇒ token usage in the last chunk only.
stream_options.include_usage	boolean	no	false	true/false	Include usage in the final chunk.
modalities	array	no	["text"]	["text","audio"] or ["text"]	Output modalities; Qwen-Omni only. ["text","audio"] enables speech output. Thinking mode forces ["text"].
audio	object	no	—	object with voice+format	Output voice + format; Qwen-Omni only, requires modalities to include "audio". Not supported in thinking mode.
audio.voice	string	Required when audio is set	—	one of the supported voice IDs for the model series (voice tables below) or a cloned-voice ID	Output voice. If omitted, the model uses the series default voice; supply it explicitly.
audio.format	string	Required when audio is set	—	wav only	Output audio container. PCM is 24000 Hz, mono, 16-bit.
temperature	float	no	model-specific	[0, 2)	Sampling temperature. Set only one of temperature/top_p.
top_p	float	no	model-specific	(0, 1.0]	Nucleus sampling threshold. Set only one of temperature/top_p.
top_k	integer	no	50 for the qwen3-omni-flash series; 20 for the other Omni models	integer ≥ 0; null or > 100 disables top_k (only top_p applies)	Candidate-token count. Non-standard → extra_body.
presence_penalty	float	no	1.5 for Qwen3.5-Omni; 0.0 for the other Omni models	[-2.0, 2.0]	Penalize already-present content. Positive ↓ repetition; negative ↑.
repetition_penalty	float	no	1.0 (= no penalty)	> 0, no strict upper limit	Consecutive-sequence repetition control. Non-standard → extra_body.
max_tokens	integer	no (Deprecated)	model’s max output length (check console)	≥ 1	Deprecated — use max_completion_tokens for new integrations. Does NOT limit chain-of-thought length. finish_reason=length when hit.
max_completion_tokens	integer	no	model’s max output length (check console)	≥ 1	Replacement for max_tokens. Limits total (chain-of-thought + final response) tokens. Recommended for reasoning models. Actual output may differ by up to ~10 tokens.
n	integer	no	1	1–4	Number of choices. Only Qwen3 non-thinking models honor n>1; if tools is passed, set n=1.
seed	integer	no	— (unset ⇒ non-reproducible)	[0, 2^31 − 1]	RNG seed for best-effort reproducibility.
stop	string | array	no	—	a string/token-id, or an array of strings or token-ids (do NOT mix string and token-id in one array)	Stop sequences; generation halts when matched.
logprobs	boolean	no	false	true/false	Return log-probabilities of output tokens. reasoning_content has no logprobs.
top_logprobs	integer	no	0	0–5	Number of candidate tokens per step. Effective only when logprobs=true.
response_format	object	no	{"type":"text"}	{"type":"text"} or {"type":"json_object"}	Output format. For json_object you must also instruct JSON in the prompt.
response_format.type	string	Required in object	text	text, json_object	—
tools	array	no	—	array of tool objects (see sub-fields)	Function-calling tool definitions. When set and the model decides to call, tool_calls is returned.
tools[].type	string	Required	—	function only	—
tools[].function.name	string	Required	—	[A-Za-z0-9_-], ≤ 64 tokens	Tool name.
tools[].function.description	string	Required	—	—	Helps the model decide when/how to call.
tools[].function.parameters	object	no	{}	valid JSON Schema	Tool input schema; empty ⇒ no params.
tool_choice	string | object	no	auto	auto, none, {"type":"function","function":{"name":"<fn>"}}	Tool-selection policy. Thinking mode cannot force a specific tool.
parallel_tool_calls	boolean	no	false	true/false	Allow parallel tool calls.
enable_thinking	boolean	no	false	true/false	Hybrid thinking. Within Omni only qwen3-omni-flash. Thinking content returns in reasoning_content; audio output is unsupported in thinking mode (set modalities=["text"]). Non-standard → extra_body.
thinking	object	no	{"type":"adaptive"}	thinking.type ∈ {adaptive,disabled}	Thinking control for MiniMax/MiniMax-M3 (documented in the same reference; not used by Qwen-Omni). Non-standard → extra_body.
thinking_budget	integer	no	model’s max chain-of-thought length (check console)	≥ 1	Max reasoning tokens (Qwen3.7/3.6/3.5/Qwen3-VL/Qwen3 incl. open-source). Non-standard → extra_body.
preserve_thinking	boolean	no	false	true/false	Feed prior assistant reasoning_content back as input (billed). Limited model list (qwen3.x-max/plus, kimi-k2.6/k2.7-code); not documented for Omni. Non-standard → extra_body.
reasoning_effort	string	no	high	high, max (low/medium→high, xhigh→max)	DeepSeek-V4 series reasoning intensity (documented in same reference; not for Omni). Non-standard → extra_body.
tool_stream	boolean	no	false	true/false (effective only when stream=true)	Stream complex tool arguments (Qwen & GLM series). For Qwen3.5/3.6/3.7 plus/flash text or all modalities. Non-standard → extra_body.
enable_code_interpreter	boolean	no	false	true/false	Enable code interpreter. Non-standard → extra_body.
enable_search	boolean	no	false	true/false	Web search. Within Omni only Qwen3.5-Omni. Non-standard → extra_body.
search_options	object	no	—	object with sub-fields below	Web-search config; effective only when enable_search=true. Non-standard → extra_body.
search_options.forced_search	boolean	no	false	true/false	Force a web search regardless of model judgment. Effective only when enable_search=true.
search_options.search_strategy	string	no	turbo	turbo, max, agent, agent_max	Search scale. For Qwen-Omni the search_strategy parameter only accepts agent.
search_options.enable_search_extension	boolean	no	false	true/false	Domain-specific search. Effective only when enable_search=true.
vl_high_resolution_images	boolean	no	false	true/false	Raise the input-image cap to the 16384-token resolution (16,777,216 px at 32×32/token for Qwen3.5-Omni); ignores max_pixels when true. Within Omni only Qwen3.5-Omni. Non-standard → extra_body.

The reference does not list a frequency_penalty parameter for this family; it lists presence_penalty and repetition_penalty.

messages[] — message types & structure

System Message object (Optional) — place at start.

• content string (Required) — system instruction.
• role string (Required) — fixed system.

User Message object (Required)

• content string | array (Required) — string for text-only; array for multi-modal input or when explicit caching is used.
• role string (Required) — fixed user.

Assistant Message object (Optional) — the model’s reply, fed back as context.

• content string (Optional) — text of the reply. Can be empty/omitted only if tool_calls is present; otherwise required. (For Qwen-Omni, assistant messages may carry text only — no audio.)
• role string (Required) — fixed assistant.
• partial boolean (Optional, default false) — enable partial (prefix-continuation) mode.
• tool_calls array (Optional) — echoed from a prior model response; each: id (Required), type=function (Required), function.name (Required), function.arguments JSON string (Required), index (Required).

Tool Message object (Optional) — tool output.

• content string (Required) — tool result serialized to a string.
• role string (Required) — fixed tool.
• tool_call_id string (Required) — id from choices[0].message.tool_calls[$index].id.

messages[].content part types (multi-modal) — full sub-fields

part type	sub-field(s)	type	required	notes / Omni limits
text	text	string	Required when type=text	Plain text.
image_url	image_url.url	string	Required when type=image_url	URL or data:image/...;base64,....
input_audio	input_audio.data	string	Required when type=input_audio	URL or data:;base64,... audio.
input_audio	input_audio.format	string	Required	e.g. mp3 or wav (input accepts AMR/WAV/3GP/3GPP/AAC/MP3).
video	video	array of image-frame URLs/base64	Required when type=video	Video-as-frame-list.
video_url	video_url.url	string	Required when type=video_url	URL or data:;base64,... of a video file (Qwen-Omni reads both visual and audio tracks).
fps	fps	float	Optional	Frames extracted per second / inter-frame interval hint. Range [0.1, 10], default 2.0. Applies to video/video_url.
min_pixels	min_pixels	integer	Optional	Lower pixel bound per image/frame; upscaled if below. Qwen3.5-Omni image input default & min = 24576 (= 24 tokens × 32×32). For video file / image-list input the default is 65536.
max_pixels	max_pixels	integer	Optional	Upper pixel bound; downscaled if above. Qwen3.5-Omni image input default 1310720 (= 1280 tokens × 32×32); the high-resolution cap is 16777216. When vl_high_resolution_images=true, max_pixels is ignored and the cap is fixed at 16777216. For video file / image-list input the default is 655360.
total_pixels	total_pixels	integer	Optional	Total-pixel budget across all extracted video frames (per-frame stays within [min_pixels,max_pixels]). Qwen3.5-Omni default & max = 184549376 (= 180224 image tokens at 32×32/token).
cache_control	cache_control.type	string (ephemeral only)	Optional (property of a user content part)	Enables explicit caching for that content.

Combined multimodal input (image+audio+video in one user message) — Qwen3.5-Omni series only. qwen3-omni-flash / qwen-omni-turbo: text + exactly one other modality, one media file per request. In multi-turn dialogs different user messages may carry different single modalities.

Input-modality limits (per series)

Audio input (input_audio): formats AMR, WAV, 3GP, 3GPP, AAC, MP3.

• Files: Qwen3.5-Omni — up to 2048 (URL) / 250 (base64); Flash & Turbo — only 1.
• Size (URL): Qwen3.5-Omni ≤ 2 GB; Flash ≤ 100 MB; Turbo ≤ 10 MB. Base64 string < 10 MB (all).
• Duration: Qwen3.5-Omni ≤ 3 h; Flash ≤ 20 min; Turbo ≤ 3 min. (Billing minimum 1 s.)

Video file input (video_url): formats MP4, AVI, MKV, MOV, FLV, WMV; visual + audio tracks billed separately.

• Files: Qwen3.5-Omni — up to 512 (URL) / 250 (base64); Flash & Turbo — only 1.
• Size (URL): Qwen3.5-Omni ≤ 2 GB; Flash ≤ 256 MB; Turbo ≤ 150 MB. Base64 < 10 MB.
• Duration: Qwen3.5-Omni ≤ 1 h; Flash ≤ 150 s; Turbo ≤ 40 s.

Image-list-as-video (video): frame counts — Qwen3.5-Omni 2–2048; Flash 2–128; Turbo 4–80 frames.

Image input (image_url): up to 2048 (URL) / 250 (base64) images; per-file ≤ 20 MB (Qwen3.5-Omni) / ≤ 10 MB (Flash, Turbo); base64 < 10 MB; both width and height must exceed 10 px; aspect ratio must not exceed 200:1.

Token/billing factors: image token density 1 token per 32×32 px (Qwen3.5-Omni & Flash) / per 28×28 px (Turbo); minimum 24 tokens per image (Qwen3.5-Omni) / 4 tokens (Flash & Turbo); default maximum 1280 tokens per image. Audio: Qwen3.5-Omni input ≈ seconds × 7 tokens, output ≈ seconds × 12.5; Flash ≈ seconds × 12.5 (in and out). Image ≈ 1 token / 32×32 px; video is split into video_tokens + audio_tokens. (Billing only — not request parameters.)

---

Response shape

Non-streaming chat.completion object (for reference; Qwen-Omni itself requires streaming)

{
  "choices": [
    { "message": { "role": "assistant", "content": "...", "reasoning_content": null,
                   "refusal": null, "audio": null, "function_call": null, "tool_calls": null },
      "finish_reason": "stop", "index": 0, "logprobs": null }
  ],
  "object": "chat.completion",
  "usage": { "prompt_tokens": 3019, "completion_tokens": 104, "total_tokens": 3123,
             "prompt_tokens_details": { "cached_tokens": 2048 } },
  "created": 1735120033, "system_fingerprint": null, "service_tier": null,
  "model": "qwen-plus", "id": "chatcmpl-..." }

In the non-stream object, message.audio is always null and completion_tokens_details.audio_tokens/reasoning_tokens are null — audio output is delivered via streaming chunks, not the non-stream body.

Streaming chat.completion.chunk objects (SSE data: lines) — the Qwen-Omni path

Text chunk:

{"id":"chatcmpl-...","choices":[{"delta":{"content":"I","function_call":null,"refusal":null,"role":"assistant","tool_calls":null},"finish_reason":null,"index":0,"logprobs":null}],"created":1757647879,"model":"qwen3.5-omni-plus","object":"chat.completion.chunk","service_tier":null,"system_fingerprint":null,"usage":null}

Audio chunk (base64 fragment in delta.audio.data):

{"choices":[{"delta":{"audio":{"data":"/v8AAAAAAA...","expires_at":1757647879,"id":"audio_..."}},"finish_reason":null,"index":0,"logprobs":null}],"object":"chat.completion.chunk","model":"qwen3.5-omni-plus","id":"chatcmpl-...","usage":null}

Thinking chunk (qwen3-omni-flash, reasoning in delta.reasoning_content):

{"choices":[{"delta":{"content":null,"role":"assistant","reasoning_content":"Hmm"},"finish_reason":null,"index":0,"logprobs":null}],"object":"chat.completion.chunk","model":"qwen3-omni-flash","id":"chatcmpl-..."}

Final / usage chunk (choices empty when include_usage=true):

{"choices":[],"object":"chat.completion.chunk",
 "usage":{"prompt_tokens":207,"completion_tokens":103,"total_tokens":310,
   "completion_tokens_details":{"audio_tokens":83,"text_tokens":20},
   "prompt_tokens_details":{"text_tokens":207}},
 "created":1757940330,"model":"qwen3.5-omni-plus","system_fingerprint":null,"service_tier":null,"id":"chatcmpl-..."}

Response fields (streaming chunk + final usage)

name	type	description
id	string	Request id; shared by all chunks of one response.
object	string	chat.completion.chunk (stream) / chat.completion (non-stream).
created	integer	Unix timestamp; shared across chunks.
model	string	Model used.
service_tier	string	Always null.
system_fingerprint	string	Always null.
choices	array	Generated choices; empty in the final chunk when include_usage=true.
choices[].index	integer	Choice index (use with n>1 to reassemble).
choices[].delta.content	string|null	Incremental text fragment.
choices[].delta.role	string	assistant; appears only in the first chunk.
choices[].delta.reasoning_content	string|null	Incremental chain-of-thought (Flash thinking mode).
choices[].delta.audio	object	Qwen-Omni audio output.
choices[].delta.audio.data	string	Incremental base64 audio (concatenate across chunks; decode as 24 kHz/mono/16-bit PCM).
choices[].delta.audio.expires_at	integer	UNIX timestamp (audio id expiry).
choices[].delta.audio.id	string	Audio segment id.
choices[].delta.audio.transcript	string	Transcript of the audio, when present.
choices[].delta.refusal	object	Always null.
choices[].delta.function_call	object	Always null (legacy; use tool_calls).
choices[].delta.tool_calls[]	array	Streamed tool calls: index, id (first chunk), type=function, function.name (first chunk), function.arguments (concatenate fragments).
choices[].finish_reason	string|null	null while generating; then stop / length / tool_calls.
choices[].logprobs	object|null	{content:[{token,bytes,logprob,top_logprobs:[{token,bytes,logprob}]}]} when logprobs=true.
usage	object	Final chunk only (with include_usage).
usage.prompt_tokens	integer	Input tokens.
usage.completion_tokens	integer	Output tokens.
usage.total_tokens	integer	Sum.
usage.completion_tokens_details.audio_tokens	integer	Output audio tokens.
usage.completion_tokens_details.text_tokens	integer	Output text tokens.
usage.completion_tokens_details.reasoning_tokens	integer	Thinking tokens (Flash).
usage.prompt_tokens_details.text_tokens	integer	Input text tokens.
usage.prompt_tokens_details.audio_tokens	integer	Input audio tokens (also where a video file’s audio-track tokens land).
usage.prompt_tokens_details.image_tokens	integer	Input image tokens.
usage.prompt_tokens_details.video_tokens	integer	Input video tokens (image list or video file).
usage.prompt_tokens_details.cached_tokens	integer	Tokens served from cache.
usage.cache_creation.ephemeral_5m_input_tokens	integer	Tokens used to create an explicit cache.
usage.cache_creation.cache_creation_input_tokens	integer	Tokens used to create an explicit cache.
usage.cache_creation.cache_type	string	ephemeral when explicit caching is used.

Non-stream message also exposes tool_calls[] (id, type=function, function.name/arguments, index), refusal (null), and function_call (null, deprecated).

---

Supported voices (audio.voice)

Model series	default voice	voice count	voice IDs
Qwen3.5-Omni (qwen3.5-omni-plus, qwen3.5-omni-flash, and their realtime variants)	Tina	55	Tina, Cindy, Liora Mira, Sunnybobi, Raymond, Ethan, Theo Calm, Serena, Harvey, Maia, Evan, Qiao, Momo, Wil, Angel, Li Cassian, Mia, Joyner, Gold, Katerina, Ryan, Jennifer, Aiden, Mione, Sunny, Dylan, Eric, Peter, Joseph Chen, Marcus, Li, Rocky, Sohee, Lenn, Ono Anna, Sonrisa, Bodega, Emilien, Andre, Radio Gol, Alek, Rizky, Roya, Arda, Hana, Dolce, Jakub, Griet, Eliška, Marina, Siiri, Ingrid, Sigga, Bea, Chloe
Qwen3-Omni-Flash — snapshot qwen3-omni-flash-2025-12-01	Cherry	49	Cherry, Serena, Ethan, Chelsie, Momo, Vivian, Moon, Maia, Kai, Nofish, Bella, Jennifer, Ryan, Katerina, Aiden, Eldric Sage, Mia, Mochi, Bellona, Vincent, Bunny, Neil, Elias, Arthur, Nini, Ebona, Seren, Pip, Stella, Bodega, Sonrisa, Alek, Dolce, Sohee, Ono Anna, Lenn, Emilien, Andre, Radio Gol, Jada, Dylan, Li, Marcus, Roy, Peter, Sunny, Eric, Rocky, Kiki
Qwen3-Omni-Flash — stable qwen3-omni-flash & snapshot qwen3-omni-flash-2025-09-15	Cherry	17	Cherry, Ethan, Nofish, Jennifer, Ryan, Katerina, Elias, Jada, Dylan, Sunny, Li, Marcus, Roy, Peter, Rocky, Kiki, Eric
Qwen-Omni-Turbo (legacy, deprecated)	Chelsie	4	Cherry, Serena, Ethan, Chelsie
qwen3-omni-30b-a3b-captioner	—	—	No audio output / no voice parameter (text-only model).

If audio.voice is omitted, the series default is used; the reference marks audio.voice Required when the audio object is supplied, so supply it explicitly.

Voice cloning

• Supported (for Qwen-Omni) only by qwen3.5-omni-plus and qwen3.5-omni-flash — not their dated snapshots.
• Mechanism: create a custom voice from a 10–20 second audio sample, with no model training required. Create via the voice-enrollment API (model: "voice-enrollment" / "qwen-voice-enrollment", action: "create_voice", set target_model, prefix, audio url) → returns a voice_id. Pass that voice_id as audio.voice (Omni). target_model set during voice creation must exactly match the synthesis model. (Audio sample requirements: WAV (16-bit) / MP3 / M4A, 10–20 s recommended, ≤ 10 MB, sample rate 16 kHz or higher.)

Audio languages

• Qwen3.5-Omni: input 113 (74 languages + 39 dialects); output 36 (29 languages + 7 dialects).
• Qwen3-Omni-Flash: input 19 (11 languages + 8 dialects); output 19 (11 + 8).
• Qwen-Omni-Turbo: input Chinese/English; output Chinese/English.

---

Examples

Minimal (text in, text+audio out)

curl -X POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions \
 -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
 -H "Content-Type: application/json" \
 -d '{
   "model": "qwen3.5-omni-plus",
   "messages": [{"role":"user","content":"Who are you?"}],
   "stream": true,
   "stream_options": {"include_usage": true},
   "modalities": ["text","audio"],
   "audio": {"voice":"Tina","format":"wav"}
 }'

Full (image + audio in, text+audio out, web search) — Qwen3.5-Omni

from openai import OpenAI
import os, base64
client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
    model="qwen3.5-omni-plus",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "https://.../dog_and_girl.jpeg"}},
            {"type": "input_audio", "input_audio": {"data": "https://.../cherry.wav", "format": "wav"}},
            {"type": "text", "text": "Describe the image and tell me what the audio is about."},
        ],
    }],
    modalities=["text", "audio"],
    audio={"voice": "Tina", "format": "wav"},
    temperature=0.7,
    top_p=0.9,
    max_completion_tokens=2000,
    stream=True,
    stream_options={"include_usage": True},
    extra_body={"enable_search": True, "search_options": {"search_strategy": "agent"}},
)
audio_b64 = ""
for chunk in completion:
    if chunk.choices:
        d = chunk.choices[0].delta
        if d.content: print(d.content, end="")
        if getattr(d, "audio", None): audio_b64 += d.audio.get("data", "")
    else:
        print(chunk.usage)
# base64.b64decode(audio_b64) -> int16 PCM @ 24000 Hz mono

Flash thinking (text-only output; no audio in thinking mode)

completion = client.chat.completions.create(
    model="qwen3-omni-flash",
    messages=[{"role": "user", "content": "Who are you?"}],
    modalities=["text"],                       # audio not allowed in thinking mode
    extra_body={"enable_thinking": True, "thinking_budget": 2048},
    stream=True,
    stream_options={"include_usage": True},
)
# reasoning streams in chunk.choices[0].delta.reasoning_content; final text in .content

Captioner (open-source, self-host) — audio in, text out only

# qwen3-omni-30b-a3b-captioner: single-turn, ONE audio input, NO text prompt, text-only output.
# Open-source (HF/ModelScope/vLLM); NOT a Model Studio OpenAI-compatible hosted endpoint.
# HF Transformers content-part type is "audio":
conversation = [{"role": "user", "content": [
    {"type": "audio", "audio": "https://.../caption2.mp3"}
]}]
# Recommended sampling (per model card / vLLM): temperature=0.6, top_p=0.95, top_k=20,
#   max_tokens=16384, max_model_len=32768. Best practice: audio <= 30 s.
# vLLM `vllm serve` exposes an OpenAI /v1/chat/completions endpoint where the part type is "audio_url":
#   {"type":"audio_url","audio_url":{"url":"https://.../caption2.mp3"}}

---

Sources

• https://www.alibabacloud.com/help/en/model-studio/qwen-omni
• https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-chat-completions
• https://www.alibabacloud.com/help/en/model-studio/omni-voice-list
• https://www.alibabacloud.com/help/en/model-studio/voice-cloning-user-guide
• https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Captioner
• https://github.com/QwenLM/Qwen3-Omni