Qwen3.8-Omni-Flash: Alibaba's Omnimodal Agent Model

The current wave of Flash-class models keeps stacking up.

We've recently covered Qwen3.8-Flash-Next, GLM-5.3-Flash, Gemini 3.8 Flash, and DeepSeek V4.1 Flash, all fighting for the cost-sensitive, fast-inference slot in the market.

Alibaba's Qwen team is now adding a new angle to that fight with Qwen3.8-Omni-Flash, a native omnimodal model built around audio and video agents.

Qwen3.8-Omni-Flash accepts text, image, audio, and video inputs with a 1M-token context window. Alibaba claims audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds it, alongside a 25%+ average improvement over its predecessor Qwen3.5-Omni-Plus across 29 evaluations. The headline agent gains are 36.5 points on WildClawBench-MM and a reported 98%+ drop in the per-hour price of audio input.

In this article, I'll cover everything new with Qwen3.8-Omni-Flash, looking at the new features, comparing it against the other Flash models we've reviewed, and walking through the benchmarks.

To learn more about how these models stack up, check out our guide to DeepSeek V4.1 Flash vs Gemini 3.8 Flash and our comparison of GLM-5.3 Flash vs Qwen3.8-Flash-Next.

TL;DR

• Qwen3.8-Omni-Flash is Alibaba's next-generation omnimodal model, replacing Qwen3.5-Omni-Plus.
• Its differentiator is audio and video agents: watching, listening, planning, and calling tools to deliver finished work like MVs and film commentary.
• It beats Gemini 3.8 Flash on most audio benchmarks and sits close behind on audio-visual reasoning.
• Reach for it if your workload is audio- or video-heavy. For pure text or code, Qwen3.8-Flash-Next scores marginally higher.
• At roughly $0.15 input / $0.47 output per 1M tokens, it undercuts most Flash-tier rivals on text pricing.

What Is Qwen3.8-Omni-Flash?

Qwen3.8-Omni-Flash is a native omnimodal model from Alibaba's Qwen team, sitting in the Flash tier as the cost-efficient, high-throughput option in the lineup.

It processes text, image, audio, and video in one model and supports a 1M-token context window, while keeping text performance comparable to a text-only model of the same size.

The goal of this generation is agent capability in audio and video, not just understanding them.

Where Qwen3.5-Omni-Plus focused on interpreting omnimodal content, Qwen3.8-Omni-Flash is built to plan tasks, call tools, and finish creative work such as video editing, music video creation, and film commentary.

Alibaba frames this as moving audio and video from perceptual inputs into the core media through which an agent reasons and executes.

The big claim here is a 36.5-point jump on WildClawBench-MM to 71.0 (Qwen3.5-Omni-Plus scored 34.5), plus a 22.3-point gain on AgenticVBench.

When tested on real meeting audio where people often talk over each other, the new system showed massive improvements over the old one:

• Figuring out who is speaking (Diarization): Errors dropped from 88% down to just 3%.
• Transcribing the actual words (cpWER): Mistakes dropped from nearly 90% down to 17%.

Qwen3.8-Omni-Flash Key Features

The features here revolve around one idea: giving an agent hands for audio and video work, not just eyes and ears.

Alibaba shipped the model alongside two tooling layers, Qwen-MM-Plugins for long-horizon workflows and Qwen-Live Harness for real-time interaction.

Audio and video agents that deliver finished work

You can hand Qwen3.8-Omni-Flash a song or a full-length film and describe what you want in one sentence, and it will plan the task, call creative tools, and return a finished video.

It supports end-to-end workflows, including music video creation, short drama translation with voice cloning, and long-form film commentary.

For a two-hour film, the model performs long-form video understanding, key-plot extraction, commentary planning, voiceover and music production, editing, and a final quality review.

It can interleave original dialogue with narration and adjust speech rate and volume so the audio layers sit together.

This matters because the fragmented transcribe-translate-dub-edit loop is exactly the coordination cost that kills these projects for solo creators.

Controllable audio-visual captioning

You define what the model looks at, how closely it looks, and how it presents the result, rather than accepting a fixed description.

Users can specify the subject, time range, level of detail, and output format for the same video.

Say you feed it a 40-minute product demo. You can ask for a 200-word overview for a listing page, then re-run the same clip asking for a JSON-schema output with per-scene timestamps, on-screen text extraction, and speaker-tagged transcripts for a downstream pipeline.

On OmniCap-IF, its content-structure and instruction-following scores (CSR and ISR) improve by 8.5 and 14.1 points over the predecessor.

Agentic long-form video understanding

Instead of processing an entire multi-hour recording, the model starts from your question and decides what to watch and listen to through coarse-to-fine evidence gathering.

This focuses compute and tokens on the relevant segments.

On OmniVideoBench, this agentic mode raises accuracy from 63.4 to 67.8 while cutting token consumption from 145,736 to 79,117 per query, a reduction of about 45.7%.

For anyone paying per token on hour-long footage, that efficiency gain is the feature.

Real-time interaction with Qwen3.8-Omni-Flash-Realtime

A separate variant, Qwen3.8-Omni-Flash-Realtime, perceives and responds while receiving live audio-visual streams and uses real-time context to call tools.

It handles speaking practice with pronunciation feedback, spatial audio perception, and knowledge injection through Skills.

The spatial audio claim is the most unusual one here: Alibaba says it is the first omnimodal model that can locate a target by sound, combining directional audio with visual information to determine the direction and distance of a source.

Reported time to first audio packet sits under 1.4 seconds even for 20-second audio-visual inputs.

Language and dialect coverage

Qwen3.8-Omni-Flash supports speech recognition across 74 languages and 39 Chinese dialects, from Mandarin and Cantonese to Sichuanese, Shanghainese, and Southern Min.

Speech generation covers 29 languages and 7 dialects.

For localization and dubbing workflows, that dialect breadth is a practical differentiator over most competing Flash models, which typically publish language counts but not dialect-level coverage.

Model development as an agent task

One of the more unusual demonstrations in the release is Alibaba using Qwen3.8-Omni-Flash to improve a smaller model rather than to serve users directly.

Given 12 hours to improve Qwen2.5-Omni-3B's Sichuan dialect recognition, the agent selected the WenetSpeech-Chuan evaluation set, built a baseline, and diagnosed errors by listening to audio samples.

Across four rounds, it created 3,413 training examples, rolled back failed attempts, and cut the character error rate from 25.79% to 15.30%, a relative reduction of about 40.7%.

It's a small experiment, but it points at a larger model driving the training of specialized small models for production.

How Does Qwen3.8-Omni-Flash Perform on the Benchmarks?

The pattern across the benchmarks is consistent: Qwen3.8-Omni-Flash wins the audio category outright, trades blows with Gemini 3.8 Flash on audio-visual tasks, and stays competitive on text against sibling models.

It is not universally ahead, and several video-reasoning rows still favor Gemini 3.8 Flash.

Audio and multi-speaker recognition

This is where the model is strongest, and where it clears Gemini 3.8 Flash by a wide margin.

On the AliMeeting multi-speaker ASR test it posts a 3.4 diarization error rate against Gemini 3.8 Flash's 72.6, and on AISHELL-4 it hits 2.8 versus Gemini's 66.4, which reflects Gemini struggling badly with overlapping speakers.

Diarization error rate measures how well a model attributes speech to the correct speaker, so a low score matters for meeting minutes, interview transcription, and any workflow with people talking over each other.

On music understanding, it also leads clearly, scoring 72.6 on MuchoMusic-RUL versus Gemini's 53.7 and 50.6 on MusTBench versus 40.3.

Audio-visual reasoning and understanding

Here, Gemini 3.8 Flash usually edges ahead, but the gap is small and closes when Qwen runs in agent mode.

• On OmniVideoBench, Qwen scores 63.4 static against Gemini's 65.2, but with the Qwen Code harness it climbs to 67.8.
• On JointAVBench it actually leads at 75.9 against Gemini's 70.4.

The video-reasoning rows are more mixed.

On Video-MME-v2, Gemini leads 71.0 to 65.0 static, and on LVOmniBench Gemini holds 70.7 to Qwen's 63.3, though Qwen jumps to 73.6 with the agent harness.

If your work is heavy video reasoning without an agent loop, Gemini still has the edge.

Multimodal tool use and agents

Agent capability is the pitch, and the WildClawBench-MM result backs it: 71.0 for Qwen3.8-Omni-Flash against 58.9 for Gemini 3.8 Flash and 41.9 for Seed 2.0 Lite.

On UniClawBench the two are basically tied at 69.6 and 69.0.

Gemini does win AgenticVBench at 45.0 to Qwen's 36.8 and OmniGAIA web search at 78.6 to 74.0.

So the tool-use story is real, but not a clean sweep. On StreamingBench, an interaction benchmark, Qwen leads decisively at 80.8 against Gemini's 79.9 and the predecessor's 57.1.

Text and coding against sibling Flash models

On pure text, Qwen3.8-Omni-Flash tracks its text-only sibling Qwen3.8-Flash-Next closely, usually within a point.

It scores 92.6 on LiveCodeBench v6 (Qwen3.8-Flash-Next: 91.9), 63.3 on SWE-bench Pro (Qwen3.8-Flash-Next: 62.5), and 91.0 on GPQA Diamond (Qwen3.8-Flash-Next: 91.7).

Against DeepSeek-V4-Flash-0731, Qwen3.8-Omni-Flash leads on SWE-bench Pro (63.3 vs 56.0) and IFBench (81.5 vs 79.2), but DeepSeek wins NL2Repo-Bench at 54.2 to 48.9.

When we tested DeepSeek-V4-Flash-0731, we found it the strongest local coding model we'd run to that point, so a Flash-tier omnimodal model matching it on repo-level software engineering is a real result.

Which Tier Should You Use?

Qwen3.8-Omni-Flash ships in two variants plus an adjustable reasoning setting, so picking the right combination depends on whether you need deep content analysis or live interaction.

The split is straightforward:

• The base Qwen3.8-Omni-Flash is built for deep understanding and creation with complete audio-visual files,
• Qwen3.8-Omni-Flash-Realtime handles continuous, low-latency streams over WebSocket or WebRTC.

You would not use the base model for a live conversation, and you would not use Realtime to render a two-hour film commentary.

On top of the variant choice, the base model exposes a reasoning_effort parameter with xhigh (default), medium, and low settings, letting you trade reasoning depth against speed and cost.

Use low for high-volume captioning where speed matters and xhigh for multi-step planning like commentary scripting.

Qwen3.8-Omni-Flash Pricing and Availability

Third-party listings put Qwen3.8-Omni-Flash at roughly $0.15 input / $0.47 output per 1M text tokens, with OmniaKey and LLM Reference both citing the official rates. Venice API lists a related Qwen 3.8 Flash route at $0.14 input / $0.49 output. These are text rates.

Audio and audio-visual input cost more, and this is the important caveat. Puter's pricing note flags that audio input carries a premium over text, and Alibaba's own comparison prices audio-visual input against Gemini 3.8 Flash rather than quoting a flat token rate. Alibaba claims the per-hour price of audio input dropped by more than 98% versus Qwen3.5-Omni-Plus and audio-visual input by more than 93%, so the premium is smaller than it was, but it is still there.

The model is generally available now on the Qianwen AI Platform and routes through Alibaba Cloud Model Studio, with Beijing and Singapore endpoints. There is no free consumer tier described in the release notes.

How to Get Access to Qwen3.8-Omni-Flash?

The API model ID is qwen3.8-omni-flash, and the realtime variant is qwen3.8-omni-flash-realtime.

It supports OpenAI-compatible Chat Completions and Responses APIs, and the multimodal tooling plugs into agent harnesses, including Codex, Claude Code, Qwen Code, and Gemini CLI through Qwen-MM-Plugins.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

completion = client.chat.completions.create(
    model="qwen3.8-omni-flash",
    messages=[{"role": "user", "content": "Summarize this meeting recording."}],
    reasoning_effort="xhigh",  # xhigh, medium, or low
)
print(completion.choices.message.content)

One reported rough edge: GitHub users have flagged frequent interruptions and session instability when running Qwen3.8-Omni-Flash with Claude Code, so budget time for harness troubleshooting.

For further info on setup, see our tutorial on running Flash-class models and our Qwen3.8-Max guide.

Final Thoughts

Alibaba is carving out a specific lane rather than chasing every Flash rival on text.

Against Qwen3.8-Flash-Next, GLM-5.3-Flash, Gemini 3.8 Flash, and DeepSeek V4.1 Flash, the pitch here is audio and video agents plus the cheapest text pricing in the group.

I'd reach for this if your work involves meeting transcription, video localization, or MV and commentary production, where the audio-visual agent tooling and the AliMeeting diarization gains genuinely stand out.

For pure text or code, Qwen3.8-Flash-Next scores marginally higher, and Gemini 3.8 Flash still leads on some video reasoning, so it is not a default swap for everyone.

If you want to get comfortable with agent workflows and multimodal models, I recommend starting with the AI Fundamentals skill track.