Alibaba launches multimodal AI model with drastically cheaper pricing

Alibaba's ↘ 4.74% Qwen team on Friday released Qwen3.8-Omni-Flash, a native omni-modal AI model that accepts text, image, audio, and video inputs through a single architecture with a 1 million token context window. The launch sent Alibaba's Hong Kong-listed shares up more than 4% in early trading.moomoo

The model, built on the Qwen3.8-Flash-Next architecture that shipped with open weights in August, marks a shift from passive multimedia understanding toward what Alibaba calls agentic capabilities — the ability to plan tasks, call external tools, and execute workflows based on audio-visual inputs.marktechpost+2

Pricing and Performance

Qwen3.8-Omni-Flash undercuts competitors on cost. API pricing for hourly audio input has dropped more than 98% compared to its predecessor, Qwen3.5-Omni-Plus, while audio-visual input costs have fallen over 93%. QwenCloud lists the model at $0.15 per million input tokens and $0.47 per million output tokens, a fraction of what Google's Alphabet Inc. ↘ 3.58% Gemini charges at comparable tiers.neowin+3

Across 29 internal benchmarks, Alibaba reported an average score improvement exceeding 25% over Qwen3.5-Omni-Plus. On OmniVideoBench, accuracy rose from 63.4 to 67.8 while token consumption dropped 45.7%, from 145,736 to 79,117 tokens. Independent benchmark results were not available at launch.marktechpost+2

Availability and Limitations

The model is live on QwenCloud, Alibaba Cloud Model Studio, and the Qwen chat interface across six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia. It supports the OpenAI-compatible API protocol, function calling, web search, and structured outputs.alibabacloud+1

Notably, Qwen3.8-Omni-Flash outputs text only — developers needing generated speech are directed to the older Qwen3.5-Omni model. No open weights were released, making self-hosting unavailable for now. Alongside the model, Alibaba open-sourced Qwen-MM-Plugins under an Apache 2.0 license, a toolkit that lets agent frameworks incorporate the model's multimodal capabilities as installable skills.marktechpost+2

The release arrives as Chinese AI firms continue to close the gap with Western rivals despite U.S. chip export restrictions, with Neowin noting that Huawei has accelerated delivery of its Ascend 960DT AI chip.neowin