Qwen3.8-Omni-Flash is Alibaba’s first omni-modal model built around agentic work rather than perception alone, and it arrived on 18 September 2026 with a 1 million token context window, text, image, audio and video inputs, and an audio pricing cut of more than 98 per cent against the previous generation. It reads a two-hour video, decides what to look at, calls tools, and returns text.
The positioning is unusually direct. Alibaba says the model approaches Google’s Gemini 3.8 Flash on audio-visual performance and beats it on overall audio performance, backed by an average improvement of more than 25 per cent across 29 benchmarks against Qwen3.5-Omni-Plus. Those are the vendor’s own figures, published with the release, and independent evaluations were not available at launch.
What makes Qwen3.8-Omni-Flash interesting is not the benchmark table, though. It is the mechanism underneath: instead of streaming a long recording from start to finish, the model works out where to look and what to listen to, then narrows down in coarse-to-fine passes. On one benchmark that raised accuracy while cutting token consumption by roughly 45.7 per cent — a rare case of a quality gain and a cost saving pointing the same way.
This article covers what the model is, the architecture it builds on, the benchmark and pricing claims with their arithmetic, the agentic perception mechanism, the open-source tooling released beside it, the self-improvement experiment Alibaba ran, and the limits that matter before anyone puts it into production.
Table of contents
- What Qwen3.8-Omni-Flash Actually Is
- The Agentic Perception Behind Qwen3.8-Omni-Flash
- The Benchmark Claims Around Qwen3.8-Omni-Flash
- What Qwen3.8-Omni-Flash Costs to Run
- The Limits Published With Qwen3.8-Omni-Flash
- The Open-Source Tooling Released Beside Qwen3.8-Omni-Flash
- The Self-Improvement Experiment Worth Noticing
- Where Qwen3.8-Omni-Flash Fits in the Market
- How Qwen3.8-Omni-Flash Compares on Price and Scope
- The Workflows Alibaba Built Qwen3.8-Omni-Flash Around
- What to Check Before Adopting Qwen3.8-Omni-Flash
- References
What Qwen3.8-Omni-Flash Actually Is
Start with the plain specification, because several details differ from what the name implies.
Inputs and outputs
The model accepts text, images, audio and video, and returns text only. Alibaba’s own documentation points developers to Qwen3.5-Omni when they need generated speech, so this is not a voice-output model despite the omni-modal label.
The context window
The context length is 1 million tokens. QwenCloud lists 991,000 maximum input tokens and 131,000 maximum output tokens, with a maximum reasoning length of 262,000 tokens.
The base architecture
Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture, the open-weight model that shipped in August 2026 previewing Alibaba’s Qwen4 design. The omni-modal variant itself is API-only: no open weights were announced at launch, so self-hosting is not an option today.
Thinking is on by default
Reasoning is enabled by default at an effort setting Alibaba labels xhigh, and can be disabled by setting reasoning effort to none. Function calling, web search, structured outputs, context caching and batch calls are all supported.
Protocols and availability
The API follows both the DashScope and OpenAI protocols, works with Chat Completions and the Responses API, and is served from six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia. It is live on QwenCloud, Alibaba Cloud Model Studio and Qwen Studio.
| Specification | Qwen3.8-Omni-Flash | Qwen3.5-Omni (for speech output) |
|---|---|---|
| Output modalities | Text only | Text and audio |
| Context window | 1M tokens | Not stated at this length |
| Audio input limit | Up to 3 hours | Up to 3 hours |
| Video input limit | 2 hours and 2 GB by URL | 1 hour |
| Audio input languages | 113 languages and dialects | 113 languages and dialects |
| Weights | API only | API only |
The Agentic Perception Behind Qwen3.8-Omni-Flash
This is the part of the release with genuine technical novelty, and it is worth understanding before the benchmark numbers.
How long video is normally handled
Most models read a long recording from beginning to end, spending compute uniformly, even when the answer occupies three minutes of a two-hour file. Every pass costs the same regardless of relevance.
What Alibaba does instead
The agent starts from the question. It decides what to watch and what to listen to, then gathers evidence over several coarse-to-fine rounds, moving from broad comprehension to detailed verification and spending tokens on the segments that matter.
The measured effect
On OmniVideoBench, which tests audio-video collaborative reasoning, applying agentic understanding raised accuracy from 63.4 to 67.8. Tokens consumed per query fell from 145,736 to 79,117 — a reduction of about 45.7 per cent.
Why that combination is unusual
Cost reductions in long-context work usually come from sampling less and accepting worse answers. Here the accuracy moved up 4.4 points while token use nearly halved, because the model is skipping irrelevant footage rather than skimming all of it.
The problem Alibaba admits is unsolved
The company notes that long audio and video are expensive to store, transmit and process across multiple inference passes, that existing agent harnesses do not natively support these modalities, and that end-to-end workflows joining omni-modal understanding to task execution remain early.
The Benchmark Claims Around Qwen3.8-Omni-Flash
All figures below come from Alibaba. No independent replication existed at publication, and the comparison model is the vendor’s own previous generation unless stated.
The headline average
Across 29 evaluations, the average score improves by more than 25 per cent over Qwen3.5-Omni-Plus. That is a composite, so it hides wide variation between tasks.
The agent benchmarks
WildClawBench-MM reaches 71.0, an improvement of 36.5 points, and UniClawBench reaches 69.6. Alibaba’s own summary puts the average agent gain across those two at 19.5 points. AgenticVBench improves by 22.3 points.
The perception benchmarks
LongAudioSpan, which measures long-form audio understanding, reaches 82.7, up 8.3 points. OmniVideoBench reaches 63.4 without agentic understanding, up 9.6 points. AliMeeting, covering Chinese multi-speaker meeting transcription, reaches 89.7.
Instruction following in captions
OmniCap-IF, which evaluates instruction-following in video captioning, reaches 28.2, with its two sub-scores improving by 8.5 and 14.1 points. The absolute number is low, which is a useful reminder of how far captioning control still has to go.
The Gemini comparison
Alibaba states that audio-visual performance approaches Gemini 3.8 Flash and that overall audio performance exceeds it. The wording is careful: parity on one axis, an advantage on the other, and no claim of superiority overall.
What Qwen3.8-Omni-Flash Costs to Run
Pricing is the most concrete part of this release, and the most consequential for anyone with a real workload.
Token pricing
QwenCloud lists $0.15 per million input tokens and $0.47 per million output tokens. Implicit cache hits cost $0.016 per million tokens, roughly a tenth of the standard input rate.
The audio reduction
Compared with Qwen3.5-Omni-Plus, the price of one hour of audio input falls by more than 98 per cent. Combined audio and video input for one hour falls by more than 93 per cent, and Alibaba’s own post puts the video-only reduction at about 89 per cent.
What a 98 per cent cut means in practice
A cost that drops by 98 per cent divides by roughly 50. Work that was economically absurd at the old rate — transcribing and summarising an entire archive, running a model over every support call — moves into the range where it is merely a budgeting question.
The cache matters more than it looks
At $0.016 per million tokens, a cache hit is about 9.4 times cheaper than a fresh input token. For repeated queries against the same long video or audio file, that ratio is where the real saving in a production workload will come from.
Comparison with the Qwen line
The sibling open-weight release Qwen3.8-Flash-Next is listed at $0.16 and $0.47 per million tokens, so the omni-modal model is priced at parity with text-only inference in the same family. That is the pricing signal worth noticing: modality no longer carries a premium.
The Limits Published With Qwen3.8-Omni-Flash
Alibaba’s documentation is unusually specific about constraints, which makes planning easier.
Media ceilings
Video files are accepted up to 2 hours and 2 GB by URL. Audio files are accepted up to 3 hours. Results are described as stable with video sampled at up to 15 frames per second.
Language coverage
Audio input supports 113 languages and dialects, matching Qwen3.5-Omni: 74 languages and 39 dialects, including Cantonese, Uyghur, Māori, Sichuanese and Shanghainese alongside the major world languages.
Spatial audio
Two-channel stereo and four-channel first-order ambisonics are supported through a multichannel flag. Alibaba claims this is the first omni-modal model with auditory localisation — estimating the direction and distance of a sound source by fusing spatial audio with what the camera sees.
Tool support in the Responses API
Function calling and web search are supported, but web search is currently the only built-in Responses tool. Anything else has to be supplied as a custom function.
What it cannot do
No speech output, no open weights, no self-hosting, and no independent benchmark verification yet. For regulated workloads, the first and last of those are the ones that usually decide the question.
The Open-Source Tooling Released Beside Qwen3.8-Omni-Flash
Because the model returns text, the surrounding tools do the media work — and Alibaba open-sourced two projects to handle that.
Qwen-MM-Plugins
Released under Apache 2.0 with the tagline “make any agent harness multimodal-native”, each capability installs as a skill plus an optional MCP server. The guided installer supports Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code and Gemini CLI.
The core plugin
The core capability lets the main model read local images and video frames natively, and visualise documents, code, data, 3D models and medical volumes for inspection, with metadata, cropping, bounding-box annotation and frame export. In its default native mode it needs no API key.
The omni capabilities
Three map directly to the launch demonstrations: omni-memory builds an audio-visual memory of a long video, omni-video2note turns a tutorial recording into an illustrated PDF, and omni-chatcut covers music-to-video, film commentary and speaker-preserving translation.
The gap in the README
Alibaba notes that most harnesses cannot yet feed audio to the main model natively, so audio is routed through the API for now. That is an honest limitation to publish on launch day.
Qwen-Live Harness
A second open-source project, built on the realtime API, is positioned as an execution environment for live interactive applications using WebSocket and WebRTC. At announcement its GitHub page returned a 404, so its status is worth rechecking before planning around it.
The Self-Improvement Experiment Worth Noticing
Buried in the launch material is an experiment that says more about Alibaba’s direction than the benchmark table does.
The task
Qwen3.8-Omni-Flash was given 12 hours to improve the Sichuanese speech recognition of a much smaller model, Qwen2.5-Omni-3B.
What it did unsupervised
It selected the evaluation set, measured the baseline, diagnosed failing audio samples and constructed training data — the loop a machine learning engineer would normally run by hand.
The result
Across four rounds it generated 3,413 data points and cut the character error rate from 25.79 per cent to 15.30 per cent, a relative reduction of about 40.7 per cent.
The division of labour it implies
Large models handle research and development; small models get deployed in operations. The agent’s role expands from calling models to improving them, which is a different economic proposition from inference alone.
The caveat
It is one task, one dialect, one vendor-run experiment, reported without an external baseline. Treat it as a direction of travel rather than a result you can plan a roadmap around.
Where Qwen3.8-Omni-Flash Fits in the Market
The competitive picture matters more than the benchmark deltas for most buyers.
Against Gemini 3.8 Flash
Alibaba’s claim is parity on audio-visual work and an advantage on audio overall, at $0.15 per million input tokens. Google’s model is the explicit target, which is itself a statement about where Alibaba thinks the comparison now sits.
Against its own open-weight line
Anyone who needs weights on their own hardware is pointed back to Qwen3.8-Flash-Next or the smaller Qwen3.8 releases. This model trades that freedom for modality coverage and a 1M-token window.
Against the long-context field
A 1 million token window is no longer rare — Tencent’s Hy4 preview shipped one as an open-weight model in August 2026 — but pairing it with native audio and video input and agentic retrieval inside the window is a narrower club.
The workflow framing
The demonstrations are all production workflows rather than capability demos: music video creation, short-drama translation for overseas release, commentary on a two-hour feature film, PDF notes generated from tutorial videos with screenshots, and extracting reusable procedures from demonstration recordings.
What the shift means
The contest over omni-modal models is moving away from recognition accuracy and towards integrating models into real workflows at a cost that survives contact with a finance team. That framing, more than any single score, is what this release is arguing.
How Qwen3.8-Omni-Flash Compares on Price and Scope
Putting the numbers beside their neighbours is the quickest way to see what is genuinely new here.
Price against the previous generation
The stated reductions are per hour of media, not per token: more than 98 per cent off an hour of audio input, more than 93 per cent off an hour of combined audio and video, and about 89 per cent off video alone. A 98 per cent cut divides the bill by roughly 50; a 93 per cent cut divides it by about 14.
Price against text-only inference
At $0.15 per million input tokens, Qwen3.8-Omni-Flash sits within a cent of the text-only Qwen3.8-Flash-Next at $0.16. Historically, sending audio or video to a model carried a premium of several multiples, and that premium has now effectively disappeared inside this family.
Scope against speech-output models
The trade-off for that price is output modality. Qwen3.5-Omni still owns speech generation, voice cloning and audio control within the Qwen line, and Alibaba’s documentation routes developers there explicitly rather than pretending the newer model covers it.
Scope against the long-context field
A 1 million token window is not unique. What is less common is a window that accepts up to three hours of audio and two hours of video natively, then applies retrieval inside it rather than forcing the caller to chunk the media first.
Where the comparison is weakest
Every figure is first-party. Nobody outside Alibaba has yet reproduced the OmniVideoBench result, the Gemini comparison or the 29-benchmark average, and the model is API-only, so independent labs cannot evaluate the weights directly.
| Cost line | Rate | Relative to standard input |
|---|---|---|
| Input tokens | $0.15 per 1M | 1.0x |
| Output tokens | $0.47 per 1M | 3.1x |
| Implicit cache hit | $0.016 per 1M | 0.11x |
| Sibling text model input | $0.16 per 1M | 1.07x |
| Audio hour vs previous generation | Over 98% lower | About 1/50th |
| Audio and video hour vs previous | Over 93% lower | About 1/14th |
The Workflows Alibaba Built Qwen3.8-Omni-Flash Around
The demonstrations tell you more about the intended buyer than the benchmark table does.
Media production
Music-to-video generation, film and video production, narration and commentary on feature-length material are all named use cases. These are jobs where a two-hour input is normal rather than exceptional.
Translation that keeps the voice
Short-form drama translation for overseas distribution is called out specifically, with the claim that translation preserves the original speaker’s vocal characteristics — the piece that usually breaks immersion in dubbed content.
Knowledge capture
Video2Note turns a tutorial recording into a PDF with screenshots, and an Omni Skill Creator extracts standard operating procedures from demonstration videos into reusable agent skills. Both are documentation tasks that nobody enjoys doing manually.
Meetings and transcription
AliMeeting, the Chinese multi-speaker meeting benchmark, reaching 89.7 is the most enterprise-relevant score in the release. Multi-speaker transcription is where most audio pipelines quietly fail.
Real-time interaction
The realtime variant supports persistent low-latency sessions over WebSocket and WebRTC, which is what any live assistant or monitoring application would need. Auditory localisation belongs to this surface rather than to batch processing.
What to Check Before Adopting Qwen3.8-Omni-Flash
A short evaluation list for teams considering it in the next quarter.
Confirm the region and data path
Six regions are supported and each requires an API key issued for that region. For anyone with data residency obligations, that is the first question, not the last.
Test agentic perception on your own footage
The 45.7 per cent token saving was measured on a benchmark. Your recordings — meeting captures, inspection video, call audio — will behave differently, and the saving is the whole economic case.
Budget with the cache, not the list price
Repeated queries over the same media are the realistic pattern, so model your costs on cache hits at $0.016 per million tokens rather than on the headline input rate.
Decide whether text output is enough
If your product needs spoken responses, this is the wrong model in the family and Alibaba says so. Pair it with Qwen3.5-Omni or a dedicated speech model instead.
Treat the numbers as vendor numbers
Every figure in this release is Alibaba’s. Until third parties reproduce the computer vision and audio results, run your own evaluation on a task you care about before committing a roadmap to Qwen3.8-Omni-Flash.
References
Qwen3.8-Omni-Flash technical announcement
Qwen3.8-Omni-Flash model page and pricing on QwenCloud
Qwen-Omni documentation, Alibaba Cloud Model Studio
Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model
Alibaba Unveils Next-Gen Omni-Modal AI with 98% Reduction in Audio Input Costs
Alibaba Releases Qwen3.8-Flash: A Multimodal 125B MoE Model That Previews Qwen4
Qwen3.8 27B: Complete Guide to the Best Open-Weight Release
Qwen3.8-Max-Preview: Alibaba Releases a 2.4 Trillion Parameter AI Model
Tencent Releases Hy4 preview: An Open-Weight MoE Model With a 1M Context Window
Gemini 3.8 Flash Is Now Available on AIxploria
DeepSeek V4.1 Flash 552B MoE Model Released on Hugging Face
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.