ByteDance

ByteDance

Runs global short-video platforms with ads

Large Language Model Inference System Engineer New Grad - Applied Machine Learning

Full-Time
No salary listed
Entry
Bachelor's, Master's
San Jose, CA, USA
In Person
Company Does Not Provide H1B Sponsorship

About the job

Requirements
  • Individuals must be completing or have recently completed a Bachelor's or Master's degree in Computer Science or a related discipline.
  • Strong command of algorithms, design patterns, and data structures, with solid knowledge of operating systems and computer architecture.
  • Proficiency in one or more programming languages such as C++ or Python, with good coding style.
  • Understanding of GPU hardware architecture, familiarity with high-performance computing software stacks such as CUDA, and experience in GPU performance analysis.
  • Strong interest in distributed systems and large-scale heterogeneous inference, with an interest in low-level principles and performance bottlenecks and in following progress in related fields.
Responsibilities
  • Participate in the engineering development of the Volcano Ark machine-learning-as-a-service inference system, optimizing large-model inference performance, cost, and stability in ultra-large-scale heterogeneous inference clusters.
  • Reduce large-model inference costs through system-level approaches such as disaggregated multi-role inference, distributed key-value cache systems, heterogeneous inference, elastic computing, and multi-tenant co-located inference.
Desired Qualifications
  • Experience in large-model inference system optimization, with practical understanding of prefill/decode disaggregation, key-value cache systems, and multi-node inference.
  • Experience in distributed network communication optimization, with deep understanding of distributed communication operator implementation and remote direct memory access principles.
  • Familiarity with resource orchestration and scheduling frameworks such as Kubernetes and Ray.

About the company

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

Debt Financing

Total Funding

$35.1B

Headquarters

Haidian, China

Founded

2012

Get referred to ByteDance

See people who can refer or advise you

Simplify Jobs

Simplify's Take

What believers are saying

  • Reuters said September 4, 2026 lenders backed a $29.6 billion ByteDance AI loan.
  • Reuters reported September 16, 2026 Anew Labs raised $290 million at $1.5 billion.
  • Feishu said H1 2026 ARR grew 2.5 times, and 90% of new customers bought AI.

What critics are saying

  • September 19, 2026 Judge Wu threatened ByteDance's $400 million settlement and 2019 consent decree.
  • September 24, 2026 U.S. copyright plaintiffs attacked ByteDance's AI training on YouTube videos.
  • If the January 2026 U.S. joint venture fails, TikTok loses access to 200 million Americans.

What makes ByteDance unique

  • September 15, 2026: ByteDance fused Doubao, Feishu, and Volcengine into one enterprise stack.
  • Feishu 8.0 lets agents join groups, draft docs, run tables, and trigger approvals.
  • Volcengine's Doubao-Seed-2.1-pro 0915 added multimodal coding and lower inference costs.

Help us improve and share your feedback! Did you find this helpful?

Benefits

Hybrid Work Options

Growth & Insights and Company News

Headcount

6 month growth

↑ 0%

1 year growth

↓ -1%

2 year growth

↑ 0%
Innovation Village
Sep 22nd, 2026
Adobe brings its Premiere video editing mobile app to Android.

Adobe brings its Premiere video editing mobile app to Android. Adobe today officially announced the expansion of its Premiere video editing mobile application to Android devices, extending its flagship video capabilities to billions of smartphone and tablet users worldwide. The long-awaited release comes nearly a year after the app made its debut on iOS and marks a strategic move by the software giant to establish a permanent footprint in the mobile-first content creation landscape. The release comes at a critical moment for mobile video producers. With short-form video dominating platforms like YouTube Shorts, Instagram Reels, and TikTok, Adobe is stepping up to compete directly against aggressive rivals such as ByteDance's CapCut and Meta's Edits. To pull creators into its creative ecosystem, Adobe is offering the core Premiere mobile app entirely for free, completely free of forced advertisements, mandatory subscriptions, or restrictive watermarks on exported projects. Desktop power engineered for touch displays. Engineered specifically for touch interfaces, the new Android app ditches the legacy codebase of older utilities like Premiere Rush to introduce a modern, high-performance editor. Despite being built for handheld devices, the core editing environment offers a professional multi-track timeline that allows creators to stack, layer, and manipulate video tracks, audio clips, sound effects, text overlays, and graphic transitions. Frame-accurate trimming, precision clip splitting, and smooth speed controls let users fine-tune their edits with desktop-style accuracy on a touch screen. The app also includes deep support for creator-made templates and royalty-free stock assets sourced directly from the Adobe Stock library, allowing beginners to kickstart projects quickly without starting from a blank canvas. Discover more Internet of Things Try CAD Software Recognizing the diversity of the Android hardware ecosystem, Adobe has optimized the application across multiple screen form factors. Beyond standard slab smartphones, the interface features tailored, adaptive layouts built for foldable devices like the Pixel Fold series and Samsung Galaxy Z Fold line, as well as large-screen Android tablets. On foldable hardware, users can seamlessly edit across both folded and unfolded screen states, keeping the timeline and video preview visible in balanced configurations without disrupting the workflow. One of the standout additions brought directly over from Adobe's pro-grade audio stack is the AI-driven "Enhance Audio" tool. Designed to address the common challenge of poor mobile microphone quality, the feature uses machine learning models to isolate speech, eliminate ambient background noise, and automatically balance vocal levels with a single tap. In addition, a direct integration partnership with YouTube allows creators to export and publish their finalized edits straight to YouTube Shorts without leaving the app interface. Alongside its traditional multi-track tools, the Premiere mobile app comes integrated with Adobe's Firefly generative AI platform. Users can take advantage of AI features such as Generative Fill to alter backgrounds or remove unwanted elements from video clips, generate custom sound effects from short text prompts, and convert static images into fluid video B-roll. While all standard video editing, multi-track processing, and exports up to 4K resolution at 60 frames per second remain completely free, Firefly AI tools rely on a credit system. Users receive a recurring monthly allocation of generative credits, with options to purchase additional credits or cloud storage through optional in-app upgrades or through an existing Creative Cloud subscription. Technical specifications and system requirements. To ensure real-time rendering and fluid playback during multi-track 4K editing, Adobe has outlined clear baseline hardware requirements for Android users. Devices must be running Android 13 or newer and possess at least 5GB of system RAM. By establishing these parameters, Adobe aims to guarantee consistent performance across supported smartphones and tablets without lag or app crashing during high-resolution processing. Adobe noted that several high-demand features remain in active development for future updates. These include: * Desktop Project Handoff: The ability to seamlessly push incomplete mobile timelines into desktop Premiere Pro. * Advanced Typography: Deep integration with Adobe Fonts. * Automated Beat Detection: Auto-snapping video cuts to background audio tracks. * Keyframe Animation & Advanced Cropping Controls. Although the launch brings feature parity with the basic iOS release, Adobe emphasized that this initial Android rollout is only the start of a broader roadmap. Several advanced features are already scheduled for upcoming software updates. High on the priority list is seamless cross-device project handoff, which will enable creators to start an edit on their phone and push the open timeline directly into Premiere Pro on desktop for advanced color grading or heavy visual effects. Other upcoming additions include full support for Adobe Fonts, keyframe animation controls, automated audio beat detection, and expanded custom captioning tools. Market implications. Industry analysts see the launch as a key driver for capturing the next generation of creative talent. With growing markets in regions like India and Latin America adopting mobile-first or mobile-only creative workflows, offering a free, uncompromised tool on the world's most widely used mobile operating system allows Adobe to build brand loyalty early. By removing high subscription barriers and annoying watermarks, Adobe is positioning Premiere Mobile as an indispensable default editor for modern creators. The Adobe Premiere app for Android is available for global download starting today on the Google Play Store.

Republic
Sep 22nd, 2026
Digital assets newsletter - september 2026.

Digital assets newsletter - september 2026. Welcome to this month's RADAR, the newsletter keeping you up to date on the latest developments in the wave of global asset tokenization. Inside this issue, Republic provide updates on SpaceX and ByteDance, Coinbase's milestone tokenized equities deployment, and share key insights from Republic CTO Brian Aznar on Solana-driven tokenization, $500 minimums, and Republic Wallet infrastructure. Enterprise & technology highlights: SpaceX & ByteDance. SpaceX: satellite scaling & dual-track compute revenue. Starlink satellites stacked on a Falcon 9 upper stage before deployment in low Earth orbit. Credit: SpaceX Following its public listing, SpaceX continues to scale its Starlink satellite constellation - surpassing 12 million active subscribers across 160 countries powered by over 7,800 active satellites in low-Earth orbit. Simultaneously, the company is monetizing its high-margin compute infrastructure. Near-term revenue is anchored by large-scale terrestrial data center leasing - including high-density GPU cluster allocations to enterprise partners like Anthropic and Google. In parallel, SpaceX is advancing its long-term Starmind orbital compute roadmap, leveraging Starship to eventually deploy solar-powered AI satellite clusters into orbit. ByteDance: enterprise maas scaling & 2026 AI capex. ByteDance continues its enterprise momentum with the launch of Doubao 2.1 Pro via Volcano Engine, driving daily token volume across enterprise and automotive partners like FAW-Volkswagen. Simultaneously, the commercial rollout of Seedance 2.0's general API has scaled ByteDance's enterprise video generation revenue to $2 billion ARR, backing a broader $23 billion 2026 AI capital investment into custom silicon and next-generation world models. The evolution of digital assets & institutional RWA rails. Coinbase deploys tokenized equities on Base. In a major institutional milestone for on-chain equities, Coinbase launched tokenized versions of major U.S. stocks - including Apple, Nvidia, Meta, and Alphabet - natively on Base under the B20 standard. Supported at launch by a coordinated ecosystem of over 50 DeFi protocols for lending, routing, and yield collateral, the rollout demonstrates a turnkey transition toward continuous, 24/7 clearing and settlement for public equities on public blockchains. Institutional RWA acceleration across yield and treasuries. On-chain real-world assets (RWAs) have experienced major institutional expansion, with total distributed tokenized asset value reaching $39.2 billion - a more than threefold increase year-over-year. Tokenized U.S. Treasuries and government money market funds remain the primary growth engine, scaling to $15.9 billion (anchored by flagship vehicles like BlackRock's $2.8 billion BUIDL), while tokenized private credit expanded to $8 billion. This rapid acceleration is driven by institutional treasury managers utilizing T+0 settlement, continuous liquidity, and yield-bearing tokens as off-exchange trading collateral. Republic ecosystem & product update. CTO Brian Aznar on RWA access, Solana infrastructure, and $500 minimums. Republic CTO Brian Aznar recently joined Solana's "The Stack" to break down how tokenization fundamentally transforms private market accessibility. Highlighting how blockchain solves mechanical inefficiencies in traditional finance, Aznar detailed Republic's deployment of daily automated loan payouts (via Profitr) and the fractionalization of institutional funds - such as Hamilton Lane's infrastructure fund - down to a $500 minimum. Aznar emphasized Solana's low, predictable gas fees, near-instant settlement, and purpose-built security token standards with integrated OFAC/AML transfer controls as critical building blocks for institutional-grade RWAs.

Pandaily
Sep 21st, 2026
ByteDance Doubao phone agent goes mcp/a2a-first with GUI fallback.

ByteDance Doubao phone agent goes mcp/a2a-first with GUI fallback. Published: September 21, 2026 Want to read in a language you're more familiar with? Doubao's consumer phone assistant prioritizes MCP and A2A protocol calls to apps, keeping GUI-agent automation as a SAEP-bounded fallback rather than the default path. ByteDance's Doubao phone assistant consumer edition is rolling out with an agent architecture that prefers structured protocol calls - MCP and A2A-style interfaces where apps expose capabilities - before falling back to GUI-agent screen control, according to September mid-month Chinese tech coverage of the consumer release and its Screen Automation Execution Protocol. The news value is the execution stack and ecosystem rules, not a handset SKU launch: Doubao is repositioning how an on-device agent reaches third-party services after earlier preview builds relied heavily on vision-driven tap-and-swipe automation. Under the reported hybrid path, when an application already exposes MCP or A2A endpoints, Doubao can pass intent and receive structured results without driving the UI. When no protocol path exists, a GUI agent remains available in Beta form, but SAEP lets developers declare page-, intent- and action-level boundaries for what screen automation may do. ByteDance set a 30-day publicity window in which automation defaults to system apps, ByteDance-owned apps and third parties that explicitly opt in; apps that refuse stay off the automation list even after the window. High-risk steps such as payments or content publishing still require user confirmation. Hands-on reports after the consumer edition shipped describe a cautious posture: the assistant may open WeChat, shopping or food-delivery apps yet decline in-app automation where SAEP or default rules block it, while ByteDance-family apps complete multi-step flows more readily. Official materials cast SAEP as an industry boundary mechanism analogous to robots.txt for agent-era clients - public, revisable declarations rather than silent scraping of every interface. For agent builders the signal is prioritization: protocol-first integration is the preferred product path, with GUI control as a constrained fallback under developer-declared limits. Whether major non-ByteDance apps adopt MCP/A2A hooks at scale will decide how often Doubao stays on the structured path versus the Beta GUI lane.

Gasgoo
Sep 20th, 2026
Roewe Jia Yue 07 debuts Doubao Cockpit Assistant.

Roewe Jia Yue 07 debuts Doubao Cockpit Assistant. Edited by Betty From Gasgoo | September 20, 2026 18:47 BJT Gasgoo Munich- SAIC Roewe announced on Sept. 17 that the all-new Roewe Jia Yue 07 features the "Doubao Cockpit Assistant." The assistant was jointly developed by SAIC Motor and ByteDance's Volcano Engine. Pre-sales for the vehicle begin on Sept. 21. The Jia Yue 07 is billed as the industry's first AI-native car. It is built on the CPP AI-native architecture. This enables deep integration between the Doubao assistant, vehicle control, navigation, driver-assistance systems, and the user ecosystem. Image Source: Roewe Auto Unlike traditional "AI-on-board" models, the Doubao Cockpit Assistant is deeply embedded in the full "perception-decision-execution" loop. It can sense hundreds of vehicle signals. The assistant supports multimodal emotion recognition and complex multi-task execution. SAIC Motor revealed that more new models will arrive by the end of the year. Relevant SAIC Audi models are also set to debut. Public records show the partnership between Roewe and the Doubao large model began in April 2025. SAIC Motor President Jia Jianxu and Volcano Engine Automotive General Manager Yang Liwei jointly announced it. That September, the Roewe M7 DMH became the first globally to feature the Doubao deep-thinking large model. This drove the daily usage rate of its voice functions to surge from 60% to 90%. Currently, the Doubao large model sees average daily token calls exceed 120 trillion. Volcano Engine has secured in-car AI partnerships with Dongfeng, Chery, Seres, Tesla China, and others, as the race for intelligent cockpit large models continues to heat up. Gasgoo not only offers timely news and profound insight about China auto industry, but also help with business connection and expansion for suppliers and purchasers via multiple channels and methods. Buyer service: [email protected] Seller Service: [email protected]

Pandaily
Sep 17th, 2026
ByteDance ships Doubao-Seed-2.1-pro 0915 with multimodal coding on Volcengine.

ByteDance ships Doubao-Seed-2.1-pro 0915 with multimodal coding on Volcengine. Published: September 17, 2026 Want to read in a language you're more familiar with? Volcengine fully released Doubao-Seed-2.1-pro 0915 with multimodal coding demos - including a ~280k-line Java ERP rebuilt from screen recordings and sketches - ~83% mergeable Luanti repo fixes, Doubao Work/TRAE integration, and 30%+ lower image/video inference token cost. On September 16, ByteDance's Volcengine said Doubao-Seed-2.1-pro updated to the 0915 release is fully available on Volcano Ark APIs, with Doubao Work upgraded in parallel and TRAE already connected. The company frames the refresh around production multimodal coding, more reliable agent delivery, and lower image and video inference token cost - reported as more than 30% below the prior generation - rather than a funding or consumer-chat story. Multimodal coding is the concrete product hook. ByteDance argues that enterprise intent often lives in screen recordings, UI sketches, and operator habits rather than clean documents. In an official case, Doubao-Seed-2.1-pro 0915 read an undocumented Java ERP of about 280,000 lines from a recording and a few sketches, inferred architecture, and generated a working mobile front end - shifting the interface from humans writing machine-readable specs toward models reading human visual expression. A second showcase used the open-source Luanti game repository - about 387,000 lines of code and 1,000 historical issues - with official patches held out. The model orchestrated multiple sub-agents for nearly 36 hours of root-cause analysis and cross-file repair; about 83% of tasks met a mergeable standard meant to reflect engineer-acceptable quality. Separately, the model rebuilt an interactive seasonal courtyard 3D scene from four design images, staging structure, water, vegetation, and camera effects. Volcengine also stresses agent reliability and cost. Prior benchmarks placed Doubao-Seed-2.1-pro in a leading cohort on Terminal Bench 2.1, SWE-Pro, SciCode, OSWorld, MobileWorld, and MMMU-Pro; the 0915 cut adds stronger evidence tracing, source authority checks, and verification for long research reports, plus denser document and engineering-drawing parsing. Token efficiency improvements for image and video inference are positioned for high-frequency enterprise API use. The release is a model-and-product update on Volcengine - not a financing announcement - and outside teams will still want to reproduce the ERP and Luanti merge rates on their own stacks.