ByteDance's Helios reaches 19.5 FPS for minute-long video generation on single GPU
ByteDance has released Helios, a 14-billion-parameter open-weight video generation model that achieves 19.5 frames per second on a single GPU while generating minute-long video clips. The researchers claim this is the first model of its scale to reach near-real-time performance at this duration. Code and model weights are publicly available.
ByteDance's Helios Reaches 19.5 FPS for Minute-Long Video Generation
ByteDance researchers have released Helios, a 14-billion-parameter open-weight video generation model capable of producing minute-long video clips at 19.5 frames per second on a single GPU.
Performance Specs
According to ByteDance, Helios is the first video model at the 14-billion-parameter scale to achieve this performance threshold. The model generates full minutes of video while maintaining near-real-time inference speeds—a significant step toward practical video generation workflows.
The 19.5 FPS performance represents a substantial improvement over existing video models, which typically require multiple GPUs or extended processing times for longer-duration content. For context, real-time video typically targets 24-30 FPS, meaning Helios approaches this threshold on consumer-grade hardware.
Open Availability
ByteDance has released both the model weights and source code publicly, enabling researchers and developers to deploy and fine-tune Helios independently. This open-weight approach contrasts with proprietary video generation services and provides a reproducible baseline for the community.
Technical Approach
While specific architectural details are not detailed in available summaries, the achievement of minute-long generation at these speeds suggests Helios employs efficient attention mechanisms or alternative computation strategies compared to earlier diffusion-based video models. The ability to run on single-GPU hardware indicates careful optimization for memory efficiency.
Context
Video generation has emerged as one of the most computationally demanding AI tasks. Models like OpenAI's Sora and competing systems typically generate shorter clips (15-60 seconds) and require significant hardware resources. ByteDance's focus on longer durations with single-GPU compatibility addresses practical deployment constraints.
The release follows ByteDance's broader investment in open-weight AI research, positioning the company alongside Meta and other organizations releasing weights and code for community advancement.
What This Means
Helios demonstrates that efficient video generation at longer durations is achievable with careful engineering. The open-weight release enables broader adoption and provides researchers with a foundation for further optimization. However, visual quality metrics—compared to proprietary systems—remain unspecified, so practical applicability depends on whether the model's output meets production standards. The 19.5 FPS figure signals that real-time video generation infrastructure is moving within reach of standard compute resources rather than requiring specialized clusters.
Related Articles
Alibaba releases Qwen 3.8, a 2.4 trillion parameter open-weight model claiming second place behind Fable 5
Alibaba has released Qwen 3.8, a 2.4 trillion parameter open-weight model that the company claims trails only Fable 5. The multimodal model processes images, videos, and documents, with a preview available through Alibaba's platforms at 10 percent of standard pricing.
Moonshot AI releases Kimi K3, largest open-weight model at 2.8 trillion parameters
Moonshot AI released Kimi K3 on July 16, 2025, an open-weight model with 2.8 trillion parameters. The model represents the largest openly available model by parameter count, entering what the industry categorizes as the 3T class.
Moonshot AI Releases Kimi K3: 2.8T Parameter Open Model at $3/$15 Per Million Tokens
Moonshot AI has released Kimi K3, a 2.8 trillion parameter model with 1 million token context window and native multimodal input. The model ranks #1 in Frontend Code Arena and #9 in Text Arena, with pricing at $3 per million input tokens and $15 per million output tokens—comparable to Claude Sonnet 5 pricing while delivering performance the company claims is near Claude Opus 4.8 and GPT-5.5.
Thinking Machines releases Inkling: 975B-parameter MoE model with Apache 2.0 license, first major US open-weight multimo
Thinking Machines Lab released Inkling, a mixture-of-experts model with 975B total parameters and 41B active parameters, trained on 45 trillion tokens across text, images, audio, and video. The Apache 2.0-licensed model supports up to 1M context and debuts alongside Inkling-Small (276B-A12B), marking what observers call the strongest US-based open-weight release to date.
Comments
Loading...