Under the January 2026 divestiture, ByteDance retained ownership of TikTok’s core recommendation algorithm intellectual property, according to Will Farrell’s prepared testimony submitted to the House Select Committee on the Strategic Competition between the United States and the Chinese Communist Party. Farrell, TikTok’s U.S. Chief Security Officer, explained that while TikTok’s U.S. operations were spun off into TikTok USDS Joint Venture LLC, ByteDance did not transfer or surrender ownership of the underlying algorithm architecture, which is built on ByteDance’s proprietary Monolith framework.

Under the January 2026 restructuring created TikTok USDS as a joint venture in which ByteDance holds a 19.9% stake, with the remainder owned by U.S. and allied investors, according to congressional letters and public disclosures.

The January 2026 restructuring created TikTok USDS as a joint venture in which ByteDance holds a 19.9% stake, with the remainder owned by U.S. and allied investors, according to congressional letters and public disclosures. Although TikTok USDS employees are responsible for reviewing the algorithm’s source code and retraining the recommendation engine on U.S. user data stored in Oracle’s American cloud infrastructure, the intellectual property rights to the algorithm’s architecture, evaluation criteria, and training methodology remain with ByteDance. Farrell’s testimony emphasized that TikTok USDS licenses the recommendation algorithm from ByteDance rather than owning it outright.

Farrell distinguished between the algorithm’s architecture and its parameters, noting that TikTok USDS retrains the model’s numerical weights exclusively on U.S.-based user data, while ByteDance retains control over the design and framework of the system. Oracle serves as a “trusted security partner,” auditing compliance with national security requirements and monitoring how U.S. data is used in the retraining process. Despite these safeguards, the retention of the algorithm’s intellectual property by ByteDance has raised concerns among lawmakers about whether the divestiture effectively removed Chinese control or influence over TikTok’s core technology.

Congressional scrutiny has focused on the potential for ByteDance, as a Chinese company subject to Chinese intelligence laws, to exert ongoing influence through its ownership of the algorithm architecture. Analysts cited in coverage of Farrell’s testimony note that Chinese law obligates companies like ByteDance to cooperate with intelligence services upon request, a legal requirement that persists despite the U.S. restructuring. This has fueled questions about whether ByteDance could modify or update the algorithm remotely, potentially affecting content prioritization on the U.S. platform.

The House Select Committee is expected to explore how U.S. user data flows within the joint venture and whether technical interfaces exist between TikTok USDS and ByteDance-controlled systems. These inquiries build on earlier congressional efforts, including letters dating back to November 2021 that demanded detailed information about TikTok’s algorithm, especially regarding its recommendations to younger users and management of harmful content. TikTok CEO Shou Zi Chew previously testified before Congress in March 2023, denying allegations that TikTok shares U.S. user data with the Chinese government, while acknowledging that Beijing-based employees had technical access to U.S. data at that time.

The January 2026 divestiture followed TikTok’s $1.5 billion “Project Texas” initiative, which aimed to segregate U.S. user data from Chinese operations by storing data exclusively in Oracle’s U.S. cloud infrastructure. Farrell’s testimony reiterated that all U.S. user data used for retraining the recommendation model is confined to systems under TikTok USDS and Oracle, with legal and technical barriers designed to prevent access by ByteDance’s China-based operations. Senator Edward Markey’s letter highlighted company pledges that TikTok USDS would conduct source code reviews to detect any “Chinese manipulation” of the algorithm.

Despite these measures, public documentation indicates that TikTok USDS has provided limited updates on the implementation timeline or technical details of these safeguards since the divestiture. The ongoing congressional investigation, including the hearing on September 15, 2026, aims to clarify the extent of ByteDance’s retained control over TikTok’s recommendation technology and the implications for U.S. national security.

ByteDance announced Seedance 2.5, its latest AI video generation model, at the Volcano Engine FORCE conference in Beijing on June 23, 2026. According to company officials, the new model can generate a single, continuous video clip up to 30 seconds long in one pass, doubling the previous 15-second limit, and is currently in global enterprise beta ahead of a planned public launch in early July.

The company emphasized that this 30-second clip is produced natively, without the need to stitch together shorter segments, a technical milestone highlighted during the Volcano Engine FORCE conference in Beijing on June 23, 2026. ByteDance representatives described the model as one of the leading commercial AI video generators in terms of clip duration at the time of announcement.

ByteDance officials said Seedance 2.5 can generate a single, continuous video clip up to 30 seconds long in one pass, doubling the maximum length of 15 seconds offered by the previous Seedance 2.0 model.

The new model supports native 4K video output and 10-bit color depth, according to ByteDance’s AI research division, which developed Seedance 2.5. These features aim to improve dynamic range and color fidelity in generated videos, making the model suitable for professional content creation and marketing workflows. Users can select aspect ratio, duration, and resolution before generation through integrated platforms, then download watermark-free clips, company materials show.

Seedance 2.5 also significantly increases the number of reference materials that can be used in a single generation, supporting up to 50 multimodal inputs, including images, video clips, audio files, and text. This represents a substantial increase from the 12-reference limit in Seedance 2.0, providing greater control and consistency over characters, products, environments, and styles across the generated clip, ByteDance officials said. Analysts and third-party reviews noted that this 50-reference capacity is among the highest available in commercial AI video models currently.

In addition to generation capabilities, Seedance 2.5 introduces localized editing features, often described as region-level or local re-draw editing. These tools allow users to modify specific areas of a video without regenerating the entire clip, enabling targeted corrections or stylistic changes within the sequence. ByteDance presentations and technical summaries highlighted improved editing control compared with Seedance 2.0, including region-level editing and 3D previsualization tools. External platforms integrating Seedance 2.5, such as PicLumen and Easemate, outlined workflows where users generate, preview, and selectively edit or regenerate segments before downloading final 4K clips.

The model is currently in a global enterprise beta phase, accessible primarily through ByteDance’s Volcano Engine platform and selected enterprise partners, company officials confirmed. Public availability is targeted for early July 2026, according to ByteDance and multiple media reports including Softonic and India Today. Until then, Seedance 2.0 remains the version accessible to most users, according to product pages and third-party guides.

Seedance 2.5 is positioned as the next iteration in ByteDance’s Seedance series, succeeding Seedance 2.0. Key improvements over the previous model include doubling the clip duration from 15 to 30 seconds, increasing the reference limit from 12 to 50, and enhancing editing and controllability tools. Analysts noted that these upgrades address common limitations in AI video workflows, such as short clip lengths, inconsistent rendering of characters and objects, and limited ability to specify context and style.

The launch of Seedance 2.5 reflects ByteDance’s broader strategy to expand its role in global AI video and content generation ecosystems. As the parent company of TikTok, ByteDance is leveraging its AI research division and the Volcano Engine platform to develop advanced video generation models aimed at professional and enterprise content workflows, including marketing, media, and creative industries. Third-party platforms like PicLumen, Easemate, and Topview have announced integration plans for Seedance 2.5, offering trial or free access for 30-second 4K video generations in anticipation of the public rollout.

Records show that Seedance 2.5’s combination of longer clip duration, higher resolution, expanded multimodal input capacity, and enhanced editing features positions it competitively among current AI video generation models. The company’s announcement at the Volcano Engine FORCE conference marks a significant step in the evolution of AI-powered video creation tools, with the public launch scheduled for early July 2026.