Alibaba’s Tongyi Lab has unveiled R1-Omni, an open-source artificial intelligence (AI) model capable of interpreting human emotions through visual and auditory data analysis.
R1-Omni utilizes Reinforcement Learning with Verifiable Reward (RLVR), enhancing its reasoning, accuracy, and adaptability. The model positions Alibaba alongside leading AI competitors like OpenAI and DeepSeek, marking a strategic advance in the AI models sector.
According to Alibaba, R1-Omni applies RLVR for the first time within a multimodal large language model. The company states: “R1-Omni is the industry’s first application of Reinforcement Learning with Verifiable Reward (RLVR) to an Omni-multimodal large language model. We focus on emotion recognition, a task where both visual and audio modalities play crucial roles, to validate the potential of combining RLVR with Omni model.”
RLVR and How it Advances R1-Omni’s Learning
RLVR operates by rewarding the AI model only when outputs meet verifiable criteria, ensuring the model’s learning process is guided by accurate and reliable feedback. For R1-Omni, this enhances its ability to accurately recognize complex emotional cues derived from both visual and auditory inputs.
The model’s training process incorporated large datasets like MAFW and DFEW, featuring over 15,000 video samples, improving its ability to generalize recognition across diverse scenarios.
To streamline training efficiency, R1-Omni integrates Group Relative Policy Optimization (GRPO), reducing reliance on traditional critic models while enabling faster comparative evaluations.
This approach is designed to accelerate the learning process without compromising performance, ensuring that R1-Omni can process complex data more efficiently and deliver consistent emotion recognition accuracy.
Alibaba’s AI Model Evolution and Competitive Focus
The introduction of R1-Omni builds on Alibaba’s broader AI model development strategy. In January 2025, Alibaba launched Qwen 2.5-Max, a Mixture-of-Experts (MoE) model designed for improved reasoning and problem-solving. The model is fully compatible with OpenAI’s API, providing developers with a seamless integration option for scalable AI applications.
February saw the release of QwQ-Max-Preview, a reasoning-focused model also based on MoE architecture. While Alibaba withheld benchmark data, it emphasized the model’s design for computational efficiency—a critical attribute given tightening U.S. restrictions on advanced AI hardware exports.
In March, Alibaba added QwQ-32B to its portfolio, offering a balance between affordability and high performance for reasoning and coding tasks. This model specifically caters to developers working within constrained computational environments, reinforcing Alibaba’s commitment to scalable and accessible AI tools.
DeepSeek’s R2 Rollout Amid Intensifying Competition
Facing Alibaba’s rapid advancements, DeepSeek expedited the release of its R2 model, initially planned for May 2025. The acceleration aimed to maintain competitive momentum amid regulatory and market pressures. DeepSeek has been navigating increased scrutiny from European regulators over GDPR compliance and facing U.S. discussions on potential restrictions tied to national security concerns.
Further complications arose from allegations that DeepSeek improperly accessed proprietary training data from OpenAI, which led to an internal investigation by Microsoft and OpenAI.
DeepSeek has also faced concerns regarding its hardware supply chain. Although it claims to have relied only on 2,048 Nvidia H800 GPUs for training its R1 reasoning model, there has been speculation about whether the company stockpiled restricted hardware prior to sanctions.
Alibaba’s Open Source and Pricing Strategy: Shaping Competitive Dynamics
Alibaba’s approach to open-source AI development plays a key role in its strategy to expand accessibility and industry influence. The release of the Wan 2.1 AI video model earlier this year underscores these efforts.
By providing open-source alternatives to proprietary platforms like OpenAI’s Sora and Google’s Veo 2, Alibaba is positioning itself as a leading advocate for accessible AI innovation.
Wan 2.1 offers features like text-to-video (T2V) and image-to-video (I2V) generation, made available under the Apache 2.0 license. This open-source approach not only lowers barriers for developers but also adds momentum to Alibaba’s long-term AI ecosystem expansion.
Complementing its open-source efforts, Alibaba’s pricing strategy further underscores its bid to capture market share. In December 2024, the company reduced prices for its Qwen-VL models by 85%, increasing affordability for businesses and developers. In December 2024, the company slashed the price of its Qwen-VL models by 85%, making its AI more accessible.
This price strategy directly challenges competitors like DeepSeek, which recently faced API access limitations due to overwhelming demand. By combining open-source access with competitive pricing, Alibaba strengthens its position among developers and enterprises, offering solutions that are both scalable and economically viable.
As Alibaba and DeepSeek continue their competition, their differing strategic approaches may shape the future trajectory of AI development in China. DeepSeek’s upcoming R2 model will be closely evaluated for its reasoning accuracy and efficiency, especially amid regulatory pressures and hardware limitations. Meanwhile, Alibaba’s emphasis on explainable AI models like R1-Omni and its open-source initiatives position it as a leader in setting new industry standards for accessibility and efficiency.
