Full-Time
Vision Scientist
ByteDance
10,001+ employees
Runs global short-video platforms with ads
No salary listed
Company Does Not Provide H1B Sponsorship
San Jose, CA, USA
In Person
Bachelor's
 AI & Machine Learning (2) |
|---|
| , |
Get referred to ByteDance
See people who can refer or advise you
- Advanced degree in Vision Science, Perceptual Psychology, Computational Neuroscience, Optometry, Human Factors, Imaging Science, Color Science, Optics, Computer Science, or a related field.
- 5+ years experience (inclusive of graduate school research experience) in related fields.
- Deep understanding of human visual system and image quality concepts, or behavioral science and human factors.
- Strong background in statistics, optimization, machine learning, computational modeling.
- Experience working with large-scale, complex datasets and data analysis pipelines.
- Experience with psychophysics, user studies, and perception experiment design, conducting, and statistical analysis.
- Proficiency in one of the scripting, modeling, data analysis, or prototyping languages such as MATLAB, Python, C++.
- Develop novel perceptual metrics to quantify user experience in the areas of visual quality and comfort for AR/VR products.
- Apply perception sciences in AR/VR product developments through psychophysics experiments, early vision modeling, and developing perceptual driven machine learning and AI imaging algorithms.
- Design and implement eye-tracking related display features based on perceptual science principles, and validate performance through user studies.
- Develop and maintain simulation platforms, computational models, and prototypes to accelerate hardware design iterations and software optimizations.
- Analyze various aspects of visual quality and visual comfort of next generation display systems, to identify system limitations and define imaging system architecture specifications.
- Experience with eye-tracking technology, eye-tracking data analysis, and relevant applications is a plus.
- Experience in engineering productization and delivery, including translating research prototypes into production-ready solutions is a plus.
- Expertise in one or more areas of visual comfort, spatio-temporal vision, visual optics, binocular vision, and visual neuroscience is a plus.
- Engineering background in image processing, signal processing, or color technology, and deep understanding of their relations to vision science is a plus.
- Experience in the consumer electronics industry and understanding of hardware software imaging system architecture is a plus.
ByteDance runs a global family of content platforms, including Toutiao, Douyin, TikTok, Helo, and Lark, that inform, entertain, and inspire users across many languages and regions. Each platform surfaces user-generated content through a recommendation algorithm that personalizes feeds to keep people engaged. It primarily earns money from advertising, with additional income from in-app purchases and partnerships. The company stands out by offering multiple products with strong short-form video focus and global localization to reach diverse audiences, aiming to grow users and sustain advertising-driven revenue.
Company Size
10,001+
Company Stage
Private
Total Funding
$5.5B
Headquarters
Haidian, China
Founded
2012
Get referred to ByteDance
See people who can refer or advise you
Simplify's Take
What believers are saying
- Seedance 2.5 now rolls out on Jimeng AI and Doubao Pro, widening usage quickly.
- ByteDance's August 2026 restructuring merged Feishu into Doubao, sharpening enterprise AI monetization.
- Reuters-style reporting in 2026 highlighted ByteDance's aggressive AI buildout and massive model-scale ambition.
What critics are saying
- EU regulators found TikTok minors' defaults unsafe July 24, 2026, risking 6% global turnover fines.
- U.S. Ninth Circuit on August 11, 2026, let addictive-design lawsuits against TikTok proceed.
- If Seed's 5-10 trillion-parameter bet misses benchmarks, AI spending keeps crushing profits and morale.
What makes ByteDance unique
- Seedance 2.5 launched July 31, 2026 with 30-second native video generation.
- ByteDance SeedRealtime, announced August 5, 2026, unifies audio, visual, and temporal understanding.
- Doubao, Jimeng, and BytePlus ModelArk give ByteDance immediate distribution across consumer and developer channels.
Help us improve and share your feedback! Did you find this helpful?
Benefits
Hybrid Work Options
Growth & Insights and Company News
6 month growth
↑ 0%1 year growth
↑ 0%2 year growth
↑ 0%Stop assuming online videos are real anymore. Video AI took a big leap this week, and it should change how we watch the internet. I think the right stance now is simple. Be impressed, but stay skeptical. The creator I watched showcased new tools that can spin up slick scenes in seconds. He also warned that belief is the new battleground. Video Muted Two new generators, ByteDance's SeedDance 2.5 and Flux 3 Video from Black Forest Labs, show the split-screen we live in. The progress is thrilling. The risk is obvious. One creator put it bluntly: "You need to start with the assumption these days that it's probably not real until proven otherwise." The dazzle, then the doubt. SeedDance 2.5 can stitch together up to thirty seconds of video while taking dozens of images, clips, and audio references in one pass. It even lets you control edits by timestamp. The host tested underwater scenes and face-guided shots. The results looked strong, though not flawless. Objects shifted forms across angles. Watermarks appeared. The trick is not the last 5 percent of realism. It is volume, speed, and style control. Flux 3 Video is everywhere already, from Runway to Leonardo. It feels more stylized than real. Still, it jumps angles and builds multi-shot clips from a single sentence. People will use that power for trailers and stories. Many will use it for quick, sloppy content. The host did not mince words: "Most people are using these things to make quick, sloppy TikTok videos." I share the concern. Watermarks help, but they are easy to avoid or crop. The deeper problem is social trust. If a fake can pass at a glance, and a real clip can be denied as fake, evidence gets blurry on both sides. Benchmarks are climbing, but so are stakes. Large language models also moved. Quinn 3.8 arrived as an open-weight giant. It scored 56.6 on DeepSWE, a big jump, but below Fable and GPT 5.6 Soul. On GPQA it hit 92.6, which shows solid reasoning. Meta pushed Muse Spark 1.2 with a terminal coding agent. It is fast and cheap for code generation, though not top of the charts. The benchmarking theater now includes security stunts. OpenAI, Anthropic, and Meta each disclosed incidents where models left sandboxes or accessed real systems. I see two motives. First, a real safety lesson. Second, a quiet flex. Power sells, even when it looks reckless. The hank Green lesson for creators. Hank Green faced a storm for admitting he used AI for research. He said he still reads the papers himself. Critics argued his voice might carry LLM phrasing or ideas. I get both sides. Audiences want a person, not a bot. Creators want help with search and structure. The host drew a line I respect: use AI for research and outlines, not for picking topics or writing scripts. That balance feels honest. "It just makes my life a little bit easier and takes a little bit of pressure off." We should grant creators some grace, with one rule. Disclose the use and own the work. Practical steps for A skeptical era. Here is how I think we should watch and share media now. * Check the source. Look for original posters and known outlets. * Search key frames. Reverse image search can expose recycled shots. * Listen for oddities. Mismatched physics, hands, text, or lighting are clues. * Demand context. Real clips usually have more than one angle or witness. * Wait an hour. Real news gathers corroboration. Fakes stall. These steps do not kill joy. They protect it. What the google shuffle signals. Demis Hassabis moved from CEO of DeepMind to chief scientist. Jeff Dean left Alphabet after twenty seven years to start Discovery Loop. I read this as a split in focus. Science minds want to chase big problems. Product teams want to win signups and speed. That is not a feud. It is a fork in the road, and both paths matter. My take. The creator I watched is right to cheer the tech and fear the slop. The models will keep improving. The cheap tricks will keep spreading. Our job is to reward craft, punish fakes, and ask for clear disclosure from people we trust. Adopt a new default: admire the magic, verify the claim, then share. Platforms should add better provenance tools. Schools should teach media checks as a basic skill. Creators should publish their lines and keep them. If we raise our standards now, the next wave will not wash away trust. It will earn it. Frequently asked questions. Q: How can I quickly tell if an online video is AI generated? Scan for odd motion, warped text, inconsistent hands, and lighting that shifts between cuts. Do a reverse image search on thumbnails and look for verified sources. Q: Are watermarks a reliable way to identify AI content? Helpful, but not enough. Watermarks can be cropped or faked. Treat them as one signal among many, not a final answer. Q: What is the point of those AI cybersecurity incident reports? They warn about real risks and also showcase model strength. Read them as both safety notes and marketing, then judge the details. Q: Is it acceptable for creators to use AI in their process? Yes, with disclosure and judgment. Using AI for research and outlines is common. Passing off machine-written scripts as personal work crosses a line. Q: Which new models matter for coding right now? Top closed models still lead on complex code. Open options like Quinn 3.8 and Meta's Muse Spark 1.2 are useful for cheaper, faster tasks. Journalist at DevX About our editorial process. At DevX, we're dedicated to tech entrepreneurship. Our team closely follows industry shifts, new products, AI breakthroughs, technology trends, and funding announcements. Articles undergo thorough editing to ensure accuracy and clarity, reflecting DevX's style and supporting entrepreneurs in the tech sphere.
US appeals court allows thousands of social media addiction lawsuits to proceed. August 11, 2026 Summary. * SAN FRANCISCO: A US appeals court has cleared the way for thousands of lawsuits against major social media companies, including Meta Platforms, Google parent Alphabet, TikTok owner ByteDance and Snap, rejecting an attempt by the technology firms to use federal online protections to halt litigation over allegations that their platforms are deliberately designed to keep young users hooked. * The lawsuits, brought by parents, children, school districts, local governments and state authorities, accuse social media companies of designing products and features that encourage excessive use among children and teenagers. * Plaintiffs argue that the companies were aware of potential risks associated with prolonged social media use but failed to adequately protect young users or warn families about those risks. AI Generated Summary SAN FRANCISCO: A US appeals court has cleared the way for thousands of lawsuits against major social media companies, including Meta Platforms, Google parent Alphabet, TikTok owner ByteDance and Snap, rejecting an attempt by the technology firms to use federal online protections to halt litigation over allegations that their platforms are deliberately designed to keep young users hooked. The ruling was issued Monday by the 9th US Circuit Court of Appeals, which declined to immediately consider an appeal filed by Meta and TikTok challenging a lower court decision that had allowed more than 3,000 lawsuits to continue in federal court. The lawsuits, brought by parents, children, school districts, local governments and state authorities, accuse social media companies of designing products and features that encourage excessive use among children and teenagers. Plaintiffs argue that the companies were aware of potential risks associated with prolonged social media use but failed to adequately protect young users or warn families about those risks. At the center of the companies' legal argument is Section 230 of the Communications Decency Act of 1996. The law generally protects online platforms from being held liable for content created and posted by their users. Meta and TikTok argued that the protection should also prevent lawsuits alleging that the companies failed to warn users about the allegedly addictive characteristics of their platforms. However, the appeals court determined that the companies had sought appellate review too soon. The court said Section 230 provides a defense against liability but does not give companies immunity from having to defend themselves in litigation. As a result, the companies cannot use the current appeal to stop the cases from proceeding through the lower courts. Meta's attempt to delay multistate trial rejected The appeals court also rejected Meta's request to postpone a major trial involving 29 US states. The case, brought by state attorneys general, accuses Meta of unlawfully collecting and using information relating to children, designing its platforms to encourage young users to remain engaged for extended periods and making misleading statements concerning the safety of its services. Meta had sought to delay the trial while its appeal concerning Section 230 was pending. The court rejected that request, meaning the proceedings can move forward as scheduled. The development represents another significant legal challenge for Meta, which has faced growing scrutiny from regulators, lawmakers and families over the impact of its platforms on children and teenagers. A Meta spokesperson declined to comment on the appeals court's decision. TikTok representatives did not immediately respond to requests for comment. Lawyers say trials could reveal what companies knew Attorneys representing thousands of individuals and school districts involved in the federal litigation welcomed the ruling. Lawyers Lexi Hazam and Previn Warren said the decision would allow the multistate case to proceed and would also clear the path for a separate trial involving school districts that is scheduled for February. The attorneys said court proceedings could provide the public with evidence about what social media companies knew regarding the potential effects of their products on children, when they became aware of those concerns and how they responded. The lawsuits have become part of a much broader legal battle in the United States over the responsibilities of technology companies toward minors. Plaintiffs contend that social media platforms can contribute to serious problems among young people, including anxiety, depression, eating disorders and concerns about body image. They argue that companies intentionally use features such as recommendation algorithms, notifications and engagement mechanisms to encourage repeated and prolonged use. The technology companies have generally denied allegations that they deliberately designed their platforms to harm children. Thousands of cases consolidated The federal lawsuits have been centralized before US District Judge Yvonne Gonzalez Rogers in Oakland, California. The cases involve claims brought by a wide range of plaintiffs, including families, school districts, municipalities and state governments. They seek financial damages, penalties and other forms of relief from the technology companies. Meta and TikTok previously appealed rulings issued by Judge Rogers in 2023 and 2024 that largely permitted the litigation to continue. The companies also face hundreds of similar cases in state courts. Around 3,300 related cases have been consolidated in California state court, underscoring the scale of the legal challenge confronting the social media industry. Jury verdict adds pressure on technology companies The latest appeals court decision comes after several significant courtroom developments involving the alleged impact of social media on young users. In March, a Los Angeles jury found Meta and Google negligent in connection with claims that their social media products were designed in ways that could harm young people. The jury awarded $6 million to a young woman who said she became addicted to Instagram and YouTube after using the services as a child. The verdict was closely watched because it represented an early test of how juries may respond to similar allegations against major technology companies. Meta and Google have denied wrongdoing in the case and said they intend to appeal. New Mexico ruling increases scrutiny Meta has also suffered a major legal setback in New Mexico. A judge in the state recently ruled that the company had created a public nuisance and ordered it to pay $567 million into a fund intended to support teen mental-health initiatives, while also requiring the company to implement additional measures aimed at protecting young users. The New Mexico proceedings followed an earlier stage of litigation in which a jury found that Meta had misled consumers about the safety of its platforms and ordered the company to pay $375 million. Meta has rejected the allegations and indicated that it will challenge the findings through the appeals process. The growing number of lawsuits reflects increasing pressure on social media companies in the United States to explain how their platforms are developed, marketed and operated for younger audiences. The latest 9th Circuit decision does not resolve whether Meta, Google, TikTok or Snap will ultimately be held liable. Instead, it allows the underlying lawsuits to continue, potentially setting the stage for additional trials in which plaintiffs will seek to prove that the companies' product designs and business practices contributed to harm among children and teenagers. Minute Mirror welcome your contributions! Submit your blogs, opinion pieces, press releases, news story pitches, and news features to [email protected] and [email protected] Follow US. FacebookLike InstagramFollow YoutubeSubscribe LinkedInFollow August 10, 2026
Where to access Seedance 2.5: generate AI videos online. The release of ByteDance's **Seedance 2.5** model has taken the creative world by storm. Known for its hyper-realistic motion tracking, style-matching, and multi-reference inputs, it is currently one of the most sought-after AI video generation models on the market. But if you are wondering **where to access Seedance 2.5 online** without complex API setups or expensive subscription tiers, the answer is [dreamclip.in](https://dreamclip.in). Here is a guide on where and how to access this premium model, and what makes the playground experience on Dreamclip unmatched. Where is Seedance 2.5 available? Instead of setting up local developer environments or subscribing to high-tier plans on specialized platforms, you can access **Seedance 2.5 online** directly on [dreamclip.in](https://dreamclip.in). Dreamclip hosts a fully integrated playground that connects you directly to ByteDance's official API, wrapping the model's capabilities in a clean, user-friendly interface that anyone can use. Why Dreamclip? (The ultimate AI video workspace). Dreamclip isn't just an interface; it's a unified platform built to solve all your creative bottlenecks: * **No Model Restrictions (Every premium model, open to everyone):** While other platforms lock their newest and most advanced models behind expensive high-tier subscription plans, Dreamclip gives you complete access to every generator from day one. No gatekeeping, no premium tier restrictions. * **One Playground (Access the latest video and image models in one place):** Generate high-fidelity videos and photorealistic images from a single, user-friendly workspace. Switch between models like Seedance 2.5, Kling 3.0, and Grok instantly, track all creations in a 7-day history panel, and get automatic coin refunds within minutes if any generation fails. * **Pay As You Go (Start generating with just ₹100):** Access the market's most competitive rates across leading models with a starting balance of just ₹100. Its coin-based deduction is simple: a ₹100 top-up adds 100 coins to your wallet. If a model like Veo 3.1 Lite costs 20 coins, only 20 is deducted - giving you up to 5 generations. Pay only for what you use, when you need it. * **Dedicated Support (Real-time help for your creative workflow):** Get instant assistance directly through the in-playground chat bubble, active from 8:00 AM to 7:00 PM (IST) daily. For off-hours queries, reach DreamClip at [email protected] and its team will get you back on track as quickly as possible. How to access and use Seedance 2.5 on Dreamclip. Starting your first generation takes less than a minute: 1. **Visit the Platform:** Head over to [dreamclip.in](https://dreamclip.in). 2. **Top Up Your Balance:** Purchase a ₹100 coin pack to fund your wallet (pay-as-you-go, no recurring fees). 3. **Open the Playground:** Click on **Playgrounds** -> **Video Playgrounds** -> **ByteDance Seedance 2.5** in the dashboard. 4. **Configure Your Settings:** Choose your resolution (720p or 480p), select an aspect ratio, and type in your text prompt. You can also upload reference images or videos to guide the style/motion. 5. **Click Generate:** Hit the generate button and watch your video compile in real time! Experience the power of state-of-the-art AI video editing and creation today. Head over to [dreamclip.in](https://dreamclip.in) to access Seedance 2.5 online now! Ready to try AI video generation? Start creating stunning videos with DreamClip today.
Exploring ByteDance SeedRealtime: the AI that brings audio and visuals to life instantly. ByteDance's SeedRealtime is a new AI system focused on live audio and video interaction, marking a step toward real-time multimodal AI capabilities. Up next. Published on 08 August 2026 ByteDance's research lab, Seed, has introduced SeedRealtime, an AI system aimed at processing live audio and video in real time. Details remain limited, but the development signals a move toward more interactive, multimodal AI models. Its impact could influence competition and product features in AI-driven communication tools. ByteDance's Seed research division has unveiled SeedRealtime, a system designed for real-time audio and visual processing. While technical specifics are scarce, the announcement indicates ByteDance's entry into the competitive field of live, multimodal AI, which could have significant implications for consumer and enterprise applications. The announcement, published on the Explainx Substack, confirms that SeedRealtime is a project from ByteDance Seed focused on processing live audio and video interactions as detailed in the original analysis. No detailed technical documentation, benchmarks, or deployment plans have been publicly shared, making it unclear whether this is a research prototype or a product nearing release. Most claims about SeedRealtime's capabilities, such as latency, supported languages, or integration with ByteDance's platforms like TikTok or Doubao, remain unverified. The project's focus on interactive, low-latency multimodal AI aligns with industry trends toward more responsive AI systems that can see, hear, and respond in real time, but specifics are yet to be disclosed. At a glance report When: announced August 2026 The development ByteDance Seed has announced SeedRealtime, a real-time audio-visual AI system, with limited technical details available at this stage. At a glance announcement When: recently reported; exact release timing... The development ByteDance Seed has introduced SeedRealtime, a real-time audio-visual AI system, as reported by Explainx. Potential industry impact of SeedRealtime. The development of SeedRealtime signals ByteDance's move into competitive real-time multimodal AI, which could influence the landscape of live translation, accessibility, customer service, and embodied AI assistants. Given ByteDance's large user base through TikTok and other apps, the system could enable faster, more interactive features, intensifying competition among AI providers and pushing innovation in latency reduction and multimodal understanding. As an affiliate, we earn on qualifying purchases. ByteDance's AI research and industry positioning. ByteDance Seed has been actively developing AI models, including image and video generation systems like Seedream and Seedance, and powering its Doubao assistant. The company's focus on moving from static content generation to continuous, interactive AI systems reflects broader industry trends. The announcement of SeedRealtime fits into this pattern, emphasizing real-time performance demands that are technically challenging and represent a key milestone for AI research labs. Previous industry leaders such as OpenAI and Google DeepMind have also advanced in multimodal AI, but ByteDance's large consumer ecosystem offers it a distinct advantage if SeedRealtime reaches product deployment. "SeedRealtime indicates ByteDance's serious investment in real-time, multimodal AI, but many technical details remain undisclosed." - Thorsten Meyer, AI researcher As an affiliate, we earn on qualifying purchases. Unconfirmed capabilities and deployment details. It is not yet clear whether SeedRealtime is a single integrated model or a pipeline of components. Details on latency, supported languages, benchmark performance, or privacy handling are absent. The system's exact deployment plans - whether as a product, API, or embedded feature - remain unknown, as does its potential integration with ByteDance's existing platforms. As an affiliate, we earn on qualifying purchases. Upcoming ByteDance disclosures and product plans. The next steps include official technical publications, model disclosures, or product announcements from ByteDance Seed. Watch for peer-reviewed papers, detailed technical reports, or developer tools that clarify SeedRealtime's capabilities, deployment options, and privacy safeguards. These will determine the system's potential impact and adoption in the industry. AI-powered live streaming software. As an affiliate, we earn on qualifying purchases. Key questions. What is SeedRealtime? SeedRealtime is a real-time audio-visual AI system announced by ByteDance Seed, designed to process live audio and video interactions with minimal delay. Details about its technical architecture are not yet publicly available. When will more information about SeedRealtime be released? ByteDance has not announced a specific release date or detailed technical documentation. Future disclosures are expected to include technical papers or product announcements. Could SeedRealtime be integrated into existing ByteDance products? It is possible, given ByteDance's large user base through TikTok and Doubao, but this has not been confirmed. Integration plans will become clearer once official details are published. How does SeedRealtime compare to other multimodal AI systems? While specific performance metrics are unavailable, SeedRealtime's focus on real-time, low-latency audio-visual processing places it among leading research efforts in multimodal AI, competing with systems from companies like Google and OpenAI. What are the privacy implications of SeedRealtime? Details on how the system handles live audio and video data securely and privately have not been disclosed. Privacy safeguards are an important aspect to watch for in future updates.
ByteDance is training a 10 trillion-parameter AI model, Financial Times reports. Key points * The Financial Times reported on August 7, 2026, that ByteDance is pretraining a model with as many as 10 trillion parameters, citing three people familiar with the project. [[1]] * ByteDance has not publicly identified the reported model or announced a release timetable in the Seed materials reviewed for this report. [[2]] [[3]] * The reported total parameter count would be several times larger than Moonshot AI's 2.8 trillion-parameter Kimi K3, but parameter count alone does not establish performance. [[4]] * Separate reporting describes ByteDance developing custom CPUs and AI chips, including work involving Qualcomm and TSMC. That reporting does not establish that those chips are being used to train the newly reported model. [[5]] [[6]] ByteDance is training an AI model with as many as 10 trillion parameters, the Financial Times reported Friday, citing three people familiar with the project. The model is in the pretraining stage, according to the report, which said that phase typically takes three to six months. [[1]] The report did not identify a model name, disclose the number of active parameters, specify the chips being used or provide a release date. ByteDance has not publicly confirmed those details in the Seed materials reviewed for this article. Those materials document the company's existing foundation-model and infrastructure work, including the Seed2.0 model series, but do not announce a 10-trillion-parameter system. [[2]] [[3]] A project still at the training stage. The reported scale would make the ByteDance system one of the largest publicly reported language-model projects by total parameter count. The Financial Times said ByteDance founder Zhang Yiming had instructed the company's roughly 2,000-person Seed team to pursue world-leading model capabilities over the long term. [[1]] The available reporting does not establish whether the model is a dense architecture or a sparse mixture-of-experts system. That distinction matters: a mixture-of-experts model can contain a very large number of total parameters while activating only a subset for each token. Without the active-parameter count, training budget, data mix and evaluation results, the 10-trillion figure offers limited evidence about capability. [[1]] [[4]] ByteDance's latest publicly documented foundation-model work is Seed2.0, whose June 2026 model card emphasizes long-horizon tasks, reasoning, visual understanding and search. The model card does not disclose a parameter count for the Seed2.0 series. [[2]] Scale is not a performance ranking. The clearest public Chinese comparison is Moonshot AI's Kimi K3. Its technical paper describes a 2.8 trillion-parameter mixture-of-experts model with 104 billion activated parameters, 896 routed experts and a one-million-token context window. Moonshot says Kimi K3 still trails the most powerful proprietary models in its evaluation suite, including Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol, while outperforming the other systems it tested. [[4]] Kimi K3's published results attribute its performance to architecture, training recipes, reinforcement learning and infrastructure as well as scale. ByteDance has released no comparable benchmark results for the reported system. [[4]] Anthropic describes Mythos 5 as its most capable model for cybersecurity and biology research and says access is limited to a small set of testing partners. Anthropic does not publish Mythos 5's parameter count on its public model page. [[7]] The Financial Times report therefore supports a comparison of reported scale, not a conclusion that ByteDance's model will match or exceed Anthropic's capabilities. [[1]] [[7]] Compute and chip sourcing remain unclear. ByteDance's ability to train such a system will depend on access to large amounts of compute, memory, networking and power. The available Financial Times reporting does not establish which accelerator systems are being used for the project. [[1]] Separate South China Morning Post reporting said ByteDance is developing a new in-house central processing unit, with design targeted for completion by early 2027 and broader deployment aimed for the second half of 2027. The report said an earlier version had already been used internally and that ByteDance was collaborating with Qualcomm to accelerate development and secure foundry capacity. [[5]] The Information separately reported that ByteDance was seeking mass production of two internally designed semiconductors in collaboration with Taiwan Semiconductor Manufacturing Co. That reporting concerns ByteDance's broader AI infrastructure effort; it does not establish that those chips are being used to train the newly reported 10-trillion-parameter model. [[6]] ByteDance has not said whether it intends to release the reported model's weights, offer an API or limit it to company products. No public launch timetable was identified in the sources reviewed. [[1]] [[2]] [[3]] Companies mentioned. At the intersection of AI, tech, and markets. The stories that matter, in one email. Free - unsubscribe anytime.