I would evaluate Qwen3.8-Omni-Flash around one question: can it complete a useful audio-video workflow with tools, at an acceptable cost? A model that summarizes footage and a system that inspects footage, finds evidence, and plans an edit need different evaluations.
Alibaba Qwen announced the model on September 18, 2026. Its published specifications describe native text, image, audio, and video input, a 1M-token context window, and text-only output. Qwen positions it as its first omni-modal model built around agentic capabilities. That positioning is more useful than treating “Flash” as the whole product description.
Start With the Deployment Contract
The output restriction is the first thing I would check. Understanding speech does not imply generating speech. The Alibaba Cloud documentation places this model in multimodal understanding and agent execution; applications requiring generated audio need an Omni variant that explicitly supports it.
The hosted model exposes Chat Completions and Responses API styles. Thinking is enabled by default, with configurable reasoning effort. Function calling, web search, implicit caching, and session caching are listed capabilities. Open weights were not announced at launch, so I would plan around hosted deployment, not assume a self-hosting path.
The context and generation limits deserve separate treatment: the context window is 1,000,000 tokens, maximum input is 991,808 tokens normally or 983,616 in thinking mode, and maximum output is 131,072 tokens. QwenCloud separately lists a reasoning allowance of up to 262K tokens. These are distinct published limits, not interchangeable budgets.
QwenCloud lists prices of $0.15 input, $0.47 output, and $0.016 implicit-cache input per 1M tokens. Limits and pricing can vary by region; I would recheck the selected endpoint before estimating production costs. A low token rate does not make repeated ingestion of long videos free.
Access and Integration Choices
I would start with the official model documentation and the hosted endpoint’s supported request format. The published API styles are useful integration options, but they are not a reason to assume every provider exposes identical media handling, tool behavior, or reasoning controls.
For an existing application comparing providers through one interface, CometAPI offers unified access to Qwen3.8-Omni-Flash. That is relevant when the evaluation spans multiple models; the technical limits still need checking against the endpoint actually used.
There is also a separate open-source component: Qwen-MM-Plugins, an Apache-2.0 toolkit for multimodal-native agent harnesses. Its workflows include video production, video-to-note conversion, reusable skills learned from demonstration videos, and audiovisual memory. The toolkit being open source does not mean the hosted model’s weights are available.
What Changes Inside a Media Workflow?
A conventional pipeline might transcribe audio, sample frames, pass those artifacts to an LLM, and leave execution to application code. Qwen3.8-Omni-Flash is designed to bring more of the understanding, planning, and tool-use loop into one agent system. External tools still matter: function calling lets the model provide structured arguments for operations outside the model.
Native audio-video reasoning is useful when neither stream is sufficient alone: associating a speaker with an on-screen person, checking whether a sound preceded an action, interpreting ambient audio, or matching spoken instructions to visible steps. These are better evaluation tasks than asking for another generic video summary.
Audio support covers 113 languages and dialects, two-channel stereo, and four-channel first-order ambisonics (FOA). Directional information matters for spatial audio grounding and multi-speaker scenes. Meetings, recorded events, and embodied-agent workflows are plausible fits, though supported input formats alone do not establish application accuracy.
Architecture and Selective Inspection
QwenCloud identifies Qwen3.8-Flash-Next as the architectural foundation. It combines Gated DeltaNet with Qwen Sparse Attention: the recurrent component compresses history into a compact state, while sparse attention accesses selected long-range context instead of applying dense attention to every token pair. That is relevant when audio and video make sequences large.
I would not copy Flash-Next’s published parameter counts into an Omni specification sheet. The exact parameter mapping for the Omni model has not been disclosed. Architectural lineage is not an exact parameter specification.
Qwen also describes active exploration of long videos: locating likely evidence, inspecting relevant moments more deeply, and choosing the next operation. That makes the surrounding harness important. Retrieval and iterative inspection are part of the evaluated system, not incidental implementation details.
Read the Benchmarks as System Results
Qwen reports an average improvement of more than 25% across 29 evaluations over Qwen3.5-Omni-Plus. The launch results are first-party measurements, not independent rankings, and had not been independently reproduced at publication.
For the tables below, static rows describe a single model run under Qwen’s launch setup. Rows marked “+ agent” include a harness, retrieval, or iterative media inspection. Prompt templates, sampling settings, media preprocessing, and harness versions matter; where those details are unavailable, I would not treat a score as independently reproducible.
Tool-Oriented Audio-Video Tasks
| Benchmark | Qwen3.8-Omni-Flash | Qwen3.5-Omni-Plus | Gemini 3.8 Flash |
|---|---|---|---|
| WildClawBench-MM | 71.0 | 34.5 | 58.9 |
| UniClawBench | 69.6 | 67.1 | 69.0 |
| AgenticVBench | 36.8 | 14.5 | 45.0 |
| OmniGAIA | 74.0 | — | 78.6 |
| StreamingBench | 80.8 | 57.1 | 79.9 |
WildClawBench-MM has the largest generational jump here, from 34.5 to 71.0. AgenticVBench improves sharply too, but Gemini 3.8 Flash remains ahead on that test and on OmniGAIA. I read this as evidence for workload-specific testing, not a universal winner.
Video Reasoning With and Without an Agent
| Benchmark | Qwen3.8-Omni-Flash | Qwen3.5-Omni-Plus | Gemini 3.8 Flash |
|---|---|---|---|
| OmniVideoBench | 63.4 | 53.8 | 65.2 |
| OmniVideoBench + Qwen Code agent | 67.8 | — | 65.2 |
| Video-MME-v2 | 65.0 | — | 71.0 |
| Video-MME-v2 + agent | 71.3 | — | 71.0 |
| LVOmniBench | 63.3 | — | 70.7 |
| LVOmniBench + agent | 73.6 | — | 70.7 |
| JointAVBench | 75.9 | — | 70.4 |
This is the distinction I would preserve in any internal evaluation report. Qwen’s static results trail Gemini on several video tests, while its agent-assisted results rise. Comparing an agent-assisted score with a static score does not isolate model quality. The gain matters for deployment only if the application supplies comparable inspection and tool capabilities.
Long Audio and Speaker Attribution
| Benchmark | Qwen3.8-Omni-Flash | Qwen3.5-Omni-Plus | Gemini 3.8 Flash |
|---|---|---|---|
| LongAudioSpan | 82.7 | 74.4 | 79.3 |
| AliMeeting DER ↓ | 3.4 | 88.1 | 72.6 |
| AliMeeting cpWER ↓ | 17.2 | 89.6 | 53.1 |
| FLEURS-ASR WER ↓ | 9.3 | 7.2 | 7.9 |
| VoiceBench | 91.6 | 92.9 | 92.3 |
Lower is better for the rows marked with a downward arrow. The AliMeeting results stand out, and LongAudioSpan improves, but FLEURS-ASR and VoiceBench regress relative to the predecessor. The evidence points toward stronger complex-audio and multi-speaker handling, not uniformly better speech recognition.
Text and coding have not disappeared from the model’s scope. Qwen reports 92.6 on LiveCodeBench v6, 63.3 on SWE-bench Pro, and 91.0 on GPQA Diamond, describing results close to the related Flash text model on several evaluations. I would validate those against the actual coding tasks and harness rather than select an Omni model on coding scores alone.
Choosing Between the Three Models
Qwen3.5-Omni-Plus remains relevant for supported speech-output workflows. It accepts text, images, audio, and video, provides function calling and web search through hosted Qwen APIs, and has smaller deployment context limits. The cited deployment does not offer an equivalent default-on reasoning mode; pricing depends on region and endpoint.
Gemini 3.8 Flash’s official specification lists 1,048,576 input tokens, text output, and text, image, video, audio, and PDF inputs. It offers low, medium, and high thinking levels, plus function calling, search grounding, code execution, and other built-in tools. It is generally available through the Gemini API and positioned for broad multimodal work, long-horizon software engineering, and agents.
Google’s listed introductory price is $0.75 input / $3.75 output per 1M tokens through December 31, 2026, versus QwenCloud’s $0.15 / $0.47. I would keep those endpoint-specific, time-bound prices separate from claims about task cost. Repeated media inspection, output length, and cache use can change the economics.
For meeting intelligence, audiovisual archives, video translation, tutorial-to-notes processing, or media-production agents, Qwen3.8-Omni-Flash belongs on the shortlist. For ordinary text chat, a dedicated text model may be simpler; for generated speech, this model’s text-only output is a hard constraint. My selection criterion would be the accuracy and cost of the complete model-and-harness workflow, with static understanding, tool execution, and repeated media processing measured separately.
Originally published at cometapi.com
Top comments (0)