Alibaba's Qwen3.8-Omni-Flash Cuts Costs 98% While Matching Gemini 3.8 Flash

Model Overview

Alibaba has released Qwen3.8-Omni-Flash , a native multimodal architecture that processes text, images, audio, and video end-to-end within a single model. The architecture incorporates Qwen3.8's Gated DeltaNet hybrid attention and sparse attention mechanisms, balancing long-text reasoning with fast generation. A real-time variant, Qwen3.8-Omni-Flash-Realtime , provides sub-second first-packet latency and supports WebSocket and WebRTC bidirectional streaming.

Pricing and Context Window

Audio input pricing drops over 98% per hour, while audio-video input pricing falls over 93% . The model natively supports a 1 million token (1M) context window .

Benchmark Results vs. Qwen3.5-Omni-Plus and Gemini 3.8 Flash

Across 29 benchmarks, the new model improves average scores by over 25% compared to Qwen3.5-Omni-Plus. Key highlights:

Agent tool calling: WildClawBench-MM jumps 36.5 points to 71.0; AgenticVBench rises 22.3 points; UniClawBench reaches 69.6.

Meeting transcription and speaker separation: AliMeeting DER and cpWER plummet from 88.11/89.61 to 3.35/17.18, cleanly handling overlapping speech.

Overall audio-video parity: Performance approaches Gemini 3.8 Flash, while pure audio understanding and music reasoning exceed Gemini 3.8 Flash.

Core Productivity Features

1. Long-Video Coarse-to-Fine Autonomous Exploration

Instead of uniformly sampling frames, the model acts as an agent: it interprets the user's goal, then actively selects which segments to watch or listen to, iterating from coarse to fine. On OmniVideoBench, this reduces token consumption by 45.7% (from 145k to 79k tokens) while accuracy improves from 63.4% to 67.8%.

2. Precise, Controllable Video Scene Description

Users define observation dimensions in the prompt:

Visual: character appearance, body motion, lighting, camera angles and shot changes.

Audio: verbatim dialogue, speaker emotion, background sound effects, ambient noise.

Alignment: millisecond-level timestamps mapping speech to on-screen speakers, output as structured JSON.

3. Large Model Bootstrapping Small Model Optimization

Alibaba demonstrated Qwen3.8-Omni-Flash acting as a "senior algorithm engineer" to autonomously improve the Sichuan dialect recognition of the edge model Qwen2.5-Omni-3B . In 12 hours, the large model selected evaluation sets, identified blind spots by listening to recordings, synthesized 3,413 targeted reinforcement samples, and ran 4 automatic iteration rounds. The small model's character error rate dropped from 25.79% to 15.30% (a 40.7% relative reduction), validating a path where large models explore and distill capabilities into deployable edge models.

Agent Workflow Ecosystem: Qwen-MM-Plugins

The release includes an open plugin library, Qwen-MM-Plugins , compatible with Claude Code, Qwen Code, and other agent environments. Five production-ready workflows are highlighted:

Short-drama one-click localization: Automatic speaker diarization, colloquial translation, voice cloning, dubbing, and mixing.

Full-length movie commentary pipeline: Plot highlight extraction, script writing, voiceover with background music, and insertion of original-dialogue highlights.

Music-to-MV generation: Beat and melody analysis, lyric timestamping, storyboard matching, and MV rendering.

Video2Note: Converts long tutorial videos into structured PDFs with key screenshots, steps, and timestamps.

Omni Skill Creator: Turns an expert's screen recording into a standardized SOP and decision tree, packaged as an Agent Skill.md file for team reuse.

Real-Time Interaction: Spatial Sound Localization

Qwen3.8-Omni-Flash-Realtime is the industry's first model supporting full-modality sound-source spatial localization . In embodied AI or home robot scenarios, it fuses binaural audio and camera video to perform 3D sound-direction estimation, pinpointing a speaker or anomalous noise in a noisy environment, then autonomously invokes obstacle avoidance, path planning, and manipulator tools to act — "hear the direction, follow the clue."

Availability and Quick Start

Both models are live on Alibaba Cloud Model Studio (Bailian) and are fully OpenAI API compatible , minimizing migration effort. A Python snippet demonstrates multimodal audio-video analysis. For terminal-based agents, the open-source Qwen-Live Harness launches real-time audio-video interaction with a single command.