Qwen3.5-Omni-Flash Realtime

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

Context / input limit

262,144 tokensSource for Context / input limit (opens in a new tab)

Maximum input tokens

196,608 tokensSource for Maximum input tokens (opens in a new tab)

Maximum output

65,536 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

Not supportedSource for Reasoning (opens in a new tab)

The provider capability table marks Thinking mode unsupported for these exact selectors.

Tool calling

SupportedSource for Tool calling (opens in a new tab)

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Structured output

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

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Streaming

SupportedSource for Streaming (opens in a new tab)

Batch processing

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

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Fine-tuning

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

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Knowledge cutoff

Not collected

Release date

2026-03-26Source for Release date (opens in a new tab)

Earliest reviewed regional platform release, not a date inferred from a snapshot ID.

Parameters

Not collected

Architecture

Thinker–Talker with Hybrid-Attention mixture of expertsSource for Architecture (opens in a new tab)

Architecture is described for the Qwen3.5-Omni family. Private Plus/Flash parameter counts are not disclosed in the reviewed sources.

Languages

View all 112 entries

• Chinese
• English
• German
• French
• Italian
• Czech
• Indonesian
• Thai
• Korean
• Polish
• Japanese
• Vietnamese
• Finnish
• Portuguese
• Spanish
• Dutch
• Russian
• Malay
• Catalan
• Swedish
• Turkish
• Ukrainian
• Romanian
• Slovak
• Danish
• Icelandic
• Norwegian (Bokmål)
• Macedonian
• Greek
• Hungarian
• Galician
• Filipino
• Croatian
• Bosnian
• Slovenian
• Bulgarian
• Kazakh
• Belarusian
• Latvian
• Estonian
• Azerbaijani
• Uyghur
• Swahili
• Hindi
• Esperanto
• Kyrgyz
• Tajik
• Cebuano
• Afrikaans
• Arabic
• Lithuanian
• Javanese
• Bengali
• Persian
• Hebrew
• Punjabi
• Gujarati
• Mongolian
• Asturian
• Kannada
• Marathi
• Interlingua
• Malayalam
• Maltese
• Norwegian Nynorsk
• Telugu
• Urdu
• Georgian
• Basque
• Tamil
• Odia
• Serbian
• Maori
• Northeastern Mandarin
• Guizhou dialect
• Cantonese
• Henan dialect
• Hong Kong Cantonese
• Shanghainese
• Shaanxi dialect
• Tianjin dialect
• Taiwanese Hokkien
• Yunnan dialect
• Anhui dialect
• Fujian dialect
• Gansu dialect
• Guangdong dialect
• Hubei dialect
• Hunan dialect
• Jiangxi dialect
• Shandong dialect
• Shanxi dialect
• Sichuan dialect
• Guangxi dialect
• Hainan dialect
• Chongqing dialect
• Changsha dialect
• Hangzhou dialect
• Hefei dialect
• Yinchuan dialect
• Zhengzhou dialect
• Shenyang dialect
• Wenzhou dialect
• Wuhan dialect
• Kunming dialect
• Taiyuan dialect
• Nanchang dialect
• Jinan dialect
• Lanzhou dialect
• Nanjing dialect
• Hakka
• Southern Min

Source for Languages (opens in a new tab)

Audio input languages and dialects explicitly listed in the guide: 112 named entries. The same guide advertises 113; its language subtotal is one higher than the enumerated list.

Context caching

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

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

API aliases

qwen3.5-omni-flash-realtime, qwen3.5-omni-flash-realtime-2026-03-15Source for API aliases (opens in a new tab)

Data retention

Standard inference prompts and responses are retained up to 30 days, subject to legal/security exceptions; stateful conversation storage and files remain until deleted. These Omni/LiveTranslate releases are not named in the current ZDR-supported roster.Source for Data retention (opens in a new tab)

Search grounding

SupportedSource for Search grounding (opens in a new tab)

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Prefix completion

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

True means supported on at least one reviewed regional endpoint; inspect access details for restrictions.

Language coverage

Provider states 113 input languages/dialects; 112 are enumerated. Output: 36 languages/dialects.Source for Language coverage (opens in a new tab)

Speech output languages

View all 36 entries

• Chinese
• English
• German
• Italian
• Portuguese
• Spanish
• Japanese
• Korean
• French
• Russian
• Thai
• Indonesian
• Arabic
• Vietnamese
• Turkish
• Finnish
• Polish
• Hindi
• Dutch
• Czech
• Urdu
• Tagalog
• Swedish
• Danish
• Hebrew
• Icelandic
• Malay
• Norwegian
• Persian
• Sichuan dialect
• Beijing dialect
• Tianjin dialect
• Nanjing dialect
• Shaanxi dialect
• Cantonese
• Southern Min

Source for Speech output languages (opens in a new tab)

Voice cloning

SupportedSource for Voice cloning (opens in a new tab)

The enrollment target_model and conversation model must match. The published target roster lists the undated API selector; pinned-selector compatibility remains unverified.

Voice reference requirements

Mono WAV (16-bit), MP3 or M4A; at least 24 kHz; maximum 60 seconds; under 10 MB; at least 3 seconds of clear continuous speech. A 10–20 second clip is recommended.Source for Voice reference requirements (opens in a new tab)

Custom voice quota

1,000 voices per account; unused voices are deleted after one yearSource for Custom voice quota (opens in a new tab)

Voice enrollment cost

USD 0.01 per successfully created voice, separate from inference. Eligible Beijing/Singapore accounts have 1,000 creations free for 90 days; deleting a voice does not restore quota.Source for Voice enrollment cost (opens in a new tab)

Stock voice catalog

56 unique named presets in the current voice reference; Tina is the defaultSource for Stock voice catalog (opens in a new tab)

The overview still says 55 voices. The imported roster follows the current per-voice reference.

Usage metering

Input audio: 7 tokens/second; output audio: 12.5 tokens/second. Accompanying text is free when speech is generated; text-only responses are charged at the text output rate.Source for Usage metering (opens in a new tab)

Request mode

Realtime API: WebSocket, WebRTC or AOQSource for Request mode (opens in a new tab)

Session duration

Up to 120 minutes per WebSocket session; connection closes at the limitSource for Session duration (opens in a new tab)

Retained audio history

Up to 80 turns and 480 cumulative secondsSource for Retained audio history (opens in a new tab)

Retained video history

Up to 50 turns and 120 cumulative secondsSource for Retained video history (opens in a new tab)

Audio formats

PCM, WAV (PCM)Source for Audio formats (opens in a new tab)

Nested session.audio input/output format fields support mono 16-bit PCM. Legacy flat fields are limited to raw PCM.

Audio sample rates

8000, 16000, 24000, 48000Source for Audio sample rates (opens in a new tab)

Nested input/output configuration: input defaults to 16 kHz, output to 24 kHz. Set format before the first audio segment; input format cannot change midstream.

Audio channels

MonoSource for Audio channels (opens in a new tab)

Image input limits

JPG/JPEG; Base64 payload ≤256 KB; up to 1080p; 1 image/second recommendedSource for Image input limits (opens in a new tab)

Tool and search restrictions

Web search and custom function calling are mutually exclusiveSource for Tool and search restrictions (opens in a new tab)

Utterance endpointing

Semantic turn detection for interruption controlSource for Utterance endpointing (opens in a new tab)

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

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

Choose a voice within Qwen3.5-Omni-Flash Realtime. These are API voice identifiers, not additional model releases. Source links document each voice.

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.

Why it is included

Inference from the official API-served release and exact Model Studio cards, checked against the reviewed official Qwen public-weight inventory. No corresponding weights are published for this named release in that inventory. Public Qwen2.5-Omni and Qwen3-Omni checkpoints are excluded; their architectures and parameter counts are not treated as the hosted model identity.

Lifecycle dates

Sources reviewed