Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Alibaba Cloud / Qwen3-Omni Flash

Qwen3-Omni Flash

Multimodal conversation with optional thinking, streamed speech and a separate text-output tariff for media inputs.

Language models (LLMs)Speech & voice
availableClosed weightsGeneral purposeVoice conversationVisionAudio understanding
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
36 of 38 specification fields collected
Audio durationUp to 20 minutes for one input audio fileSee specification notes below
File size limit100 MB for a public URL; encoded Base64 string below 10 MBSee specification notes below
Input / 1M tokens$0.43Alibaba Model Studio · Singapore
Output / 1M tokens$1.66USD · observed 14 Oct 2026

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
65,536 tokensSource for Context / input limit (opens in a new tab)
Maximum input tokens
49,152 tokensSource for Maximum input tokens (opens in a new tab)
Maximum output
16,384 tokensSource for Maximum output (opens in a new tab)
Input modalities
Text, Image, Audio, VideoSource for Input modalities (opens in a new tab)
Output modalities
Text, AudioSource for Output modalities (opens in a new tab)
Reasoning
SupportedSource for Reasoning (opens in a new tab)
Tool calling
SupportedSource for Tool calling (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Structured output
Not supportedSource for Structured output (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Streaming
SupportedSource for Streaming (opens in a new tab)
Batch processing
Not supportedSource for Batch processing (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Fine-tuning
Not supportedSource for Fine-tuning (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Knowledge cutoff
Not collected
Release date
2025-12-04Source for Release date (opens in a new tab)

Earliest reviewed regional platform release, not a date inferred from the API identifier.

Parameters
Not collected
Architecture
Mixture-of-experts Thinker–TalkerSource for Architecture (opens in a new tab)

Family architecture only. Private Flash parameter counts are not established.

Languages
View all 19 entries
  • Chinese
  • English
  • German
  • French
  • Italian
  • Thai
  • Korean
  • Japanese
  • Russian
  • Spanish
  • Portuguese
  • Sichuan dialect
  • Shanghainese
  • Cantonese
  • Southern Min
  • Shaanxi dialect
  • Nanjing dialect
  • Tianjin dialect
  • Beijing dialect
Source for Languages (opens in a new tab)

Named speech input languages/dialects; this is not the complete text-language roster.

Context caching
Not supportedSource for Context caching (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

API aliases
qwen3-omni-flashSource for API aliases (opens in a new tab)

Only this currently listed API selector is assigned to the record. Retired rolling/snapshot mappings are reviewed separately.

Request quotas
60 RPM and 100,000 TPM in Beijing and SingaporeSource for Request quotas (opens in a new tab)
Stock voice catalog
17 named presets; Cherry is the defaultSource for Stock voice catalog (opens in a new tab)
Usage metering
Audio input/output: 12.5 tokens/second, with subsecond clips billed as one second.Source for Usage metering (opens in a new tab)
Output billing
Text-only responses and media-conditioned text responses have different tariffs. For speech output, accompanying text is not charged.Source for Output billing (opens in a new tab)
Data retention
Standard inference prompts/responses may be retained for up to 30 days with legal/security exceptions; stateful records/files remain until deleted. These selectors are not listed in the reviewed zero-data-retention roster.Source for Data retention (opens in a new tab)
Maximum input in thinking mode
16,384Source for Maximum input in thinking mode (opens in a new tab)
Maximum output tokens while thinking
16,384Source for Maximum output tokens while thinking (opens in a new tab)
Maximum thinking tokens
32,768Source for Maximum thinking tokens (opens in a new tab)
Speech output languages
View all 19 entries
  • Chinese
  • English
  • German
  • French
  • Italian
  • Thai
  • Korean
  • Japanese
  • Russian
  • Spanish
  • Portuguese
  • Sichuan dialect
  • Shanghainese
  • Cantonese
  • Southern Min
  • Shaanxi dialect
  • Nanjing dialect
  • Tianjin dialect
  • Beijing dialect
Source for Speech output languages (opens in a new tab)

Guide lists Thai, but the current per-voice table does not explicitly list a Thai preset. Check voice-specific language support.

Language coverage
Model card claims 119 text languages and 20 speech languagesSource for Language coverage (opens in a new tab)

The current speech guide enumerates 19 languages/dialects; their named entries are retained separately.

Request mode
Streaming Chat Completions, with optional enable_thinkingSource for Request mode (opens in a new tab)
Reasoning controls
Off by default; enable_thinking=true enables reasoning and disables audio outputSource for Reasoning controls (opens in a new tab)
Audio duration limits
Up to 20 minutes for one input audio fileSource for Audio duration limits (opens in a new tab)
Maximum audio file
100 MB for a public URL; encoded Base64 string below 10 MBSource for Maximum audio file (opens in a new tab)
Video input duration
Up to 150 seconds for one input video fileSource for Video input duration (opens in a new tab)
Source video limits
256 MB for a public URL; encoded Base64 string below 10 MB; image-list video accepts 2–128 framesSource for Source video limits (opens in a new tab)
Image input limits
Up to 10 MB per image; encoded Base64 string below 10 MBSource for Image input limits (opens in a new tab)
Image token calculation
Minimum 4 and default maximum 1,280 tokens; the newer high-resolution parameter is unsupportedSource for Image token calculation (opens in a new tab)
Search grounding
Not supportedSource for Search grounding (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Prefix completion
Not supportedSource for Prefix completion (opens in a new tab)

Supported means at least one reviewed regional endpoint; inspect access details for regional differences.

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Alibaba Model Studio · Singapore

Provider source (opens in a new tab)
API model ID
qwen3-omni-flash
Access
https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions. Use a matching regional workspace API key.
Regions
Singapore
Streaming
SupportedSource for Streaming (opens in a new tab)
Reasoning
SupportedSource for Reasoning (opens in a new tab)
Request quotas
60 RPM; 100,000 TPMSource for Request quotas (opens in a new tab)
Tool calling
Not supportedSource for Tool calling (opens in a new tab)
Structured output
Not supportedSource for Structured output (opens in a new tab)
Search grounding
Not supportedSource for Search grounding (opens in a new tab)
Prefix completion
Not supportedSource for Prefix completion (opens in a new tab)
Context caching
Not supportedSource for Context caching (opens in a new tab)
Batch processing
Not supportedSource for Batch processing (opens in a new tab)
Fine-tuning
Not supportedSource for Fine-tuning (opens in a new tab)
INPUT / 1M TOKENS$0.43
OUTPUT / 1M TOKENS$1.66

Non-thinking, text-only input and text-only output. Media input changes the text-output tariff. Speech output and thinking-mode limits must be evaluated separately. Cached input price: not collected.

USD · observed 14 Oct 2026 · Pricing evidence (opens in a new tab)
Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
MultimodalInput: Text1$0.431M tokens
MultimodalInput: Audio1$3.811M tokens
MultimodalInput: Vision1$0.781M tokens
MultimodalOutput: Text (When input contains only text)1$1.661M tokens
MultimodalOutput: Text (When input contains images/audio/video)1$3.061M tokens
MultimodalOutput: Text&Audio (Output text is not charged)1$15.111M tokens
Multimodal thinkingInput: Text(Thinking)1$0.431M tokens
Multimodal thinkingInput: Audio(Thinking)1$3.811M tokens
Multimodal thinkingInput: Vision(Thinking)1$0.781M tokens
Multimodal thinkingOutput: Text (in thinking mode, when input contains only text)1$1.661M tokens
Multimodal thinkingOutput: Text (in thinking mode, when the input contains images/audio/video)1$3.061M tokens

Rate conditions

  1. USD per million tokens; visual and audio input are separate charges. Text-only responses use a different output rate from responses to media input. Speech output bills audio only, not accompanying text. Thinking uses the separately listed mode rows and does not generate speech. The directory estimate uses non-thinking pure text.
Pricing source (opens in a new tab)

Data & enterprise terms

Standard inference prompts/responses may be retained for up to 30 days with legal/security exceptions; stateful records/files remain until deleted. These selectors are not listed in the reviewed zero-data-retention roster.

Read the service policy (opens in a new tab)

Alibaba Model Studio · China (Beijing)

Provider source (opens in a new tab)
API model ID
qwen3-omni-flash
Access
https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-mode/v1/chat/completions. Use a matching regional workspace API key.
Regions
China (Beijing)
Streaming
SupportedSource for Streaming (opens in a new tab)
Reasoning
SupportedSource for Reasoning (opens in a new tab)
Request quotas
60 RPM; 100,000 TPMSource for Request quotas (opens in a new tab)
Tool calling
SupportedSource for Tool calling (opens in a new tab)
Structured output
Not supportedSource for Structured output (opens in a new tab)
Search grounding
Not supportedSource for Search grounding (opens in a new tab)
Prefix completion
Not supportedSource for Prefix completion (opens in a new tab)
Context caching
Not supportedSource for Context caching (opens in a new tab)
Batch processing
Not supportedSource for Batch processing (opens in a new tab)
Fine-tuning
Not supportedSource for Fine-tuning (opens in a new tab)
INPUT / 1M TOKENS$0.258
OUTPUT / 1M TOKENS$0.989

Non-thinking, text-only input and text-only output. Media input changes the text-output tariff. Speech output and thinking-mode limits must be evaluated separately. Cached input price: not collected.

USD · observed 14 Oct 2026 · Pricing evidence (opens in a new tab)
Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
MultimodalInput: Text1$0.2581M tokens
MultimodalInput: Audio1$2.2651M tokens
MultimodalInput: Vision1$0.4731M tokens
MultimodalOutput: Text (When input contains only text)1$0.9891M tokens
MultimodalOutput: Text (When input contains images/audio/video)1$1.8211M tokens
MultimodalOutput: Text&Audio (Output text is not charged)1$8.9741M tokens
Multimodal thinkingInput: Text(Thinking)1$0.2581M tokens
Multimodal thinkingInput: Audio(Thinking)1$2.2651M tokens
Multimodal thinkingInput: Vision(Thinking)1$0.4731M tokens
Multimodal thinkingOutput: Text (in thinking mode, when input contains only text)1$0.9891M tokens
Multimodal thinkingOutput: Text (in thinking mode, when the input contains images/audio/video)1$1.8211M tokens

Rate conditions

  1. USD per million tokens; visual and audio input are separate charges. Text-only responses use a different output rate from responses to media input. Speech output bills audio only, not accompanying text. Thinking uses the separately listed mode rows and does not generate speech. The directory estimate uses non-thinking pure text.
Pricing source (opens in a new tab)

Data & enterprise terms

Standard inference prompts/responses may be retained for up to 30 days with legal/security exceptions; stateful records/files remain until deleted. These selectors are not listed in the reviewed zero-data-retention roster.

Read the service policy (opens in a new tab)

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Voice choices

17 voices

Choose a voice within Qwen3-Omni Flash. These are API voice identifiers, not additional model releases. Source links document each voice.

Published voice catalog
Voice / API IDLanguage & accentDelivery
CherryCherryChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
EthanEthanChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
NofishNofishChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
JenniferJenniferChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
RyanRyanChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
KaterinaKaterinaChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
EliasEliasChinese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, KoreanProvider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)
Shanghai - JadaJadaShanghainese, English, French, German, Russian, Italian, Spanish, Portuguese, Japanese, Korean
Shanghai		Provider preset voice. Language availability is listed for each voice.Voice source (opens in a new tab)

What the measurements say

No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Inference from the authors explicitly distinguishing their in-house Flash variants from the public Qwen3-Omni-30B-A3B checkpoints, checked against the official Qwen weight inventory and exact API listings. Public checkpoint parameter counts, licenses and benchmark results are not assigned to this hosted release.

Classification evidence (opens in a new tab)

Lifecycle dates

Platform release · Singapore4 Dec 2025Source (opens in a new tab)
Platform release · China (Beijing)4 Dec 2025Source (opens in a new tab)

Sources reviewed