News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgvidSeed1.5-VL

Seed1.5-VL

Multimodal vision-language model with reasoning capabilities

Seed1.5-VL is a vision-language foundation model released in May 2025 by a large team of researchers. It combines a compact vision encoder with a Mixture-of-Experts language model to handle multimodal understanding and reasoning tasks effectively.

Share on WhatsApp

  • Uses 20 billion active parameters in MoE architecture
  • Top performance on 38 of 60 public benchmarks
  • Outperforms rivals in GUI control and gameplay
  • Includes 532M-parameter vision encoder
  • Strong ability in visual puzzle reasoning

What it is

Seed1.5-VL is a foundation model designed for general-purpose multimodal understanding and reasoning. Released in May 2025, it integrates a vision encoder with 532 million parameters and a language model using a Mixture-of-Experts architecture. The language component activates 20 billion parameters during operation. This design allows the system to process text, images, and video while maintaining a relatively compact structure compared to larger dense models. It aims to support diverse applications requiring visual analysis and logical deduction.

How it came about

The model was developed by a large consortium of researchers, with Dong Guo listed as the first author among nearly two hundred contributors. The technical report detailing its construction appeared on arXiv in May 2025. The team documented their experiences in model design, data construction, and training stages to share insights with the broader research community. The release included access via the Volcano Engine platform under a specific model identifier.

Inside the model

The architecture pairs a dedicated vision encoder with a Mixture-of-Experts large language model. A Mixture-of-Experts system uses multiple sub-networks but activates only a subset for each input, improving efficiency. Here, the active parameter count stands at 20 billion, while the vision encoder contains 532 million parameters. Training involved extensive data construction and multi-stage processes. The designers focused on enabling strong performance across various benchmarks despite the compact size relative to some competitors.

What it can do

The model achieves state-of-the-art results on 38 out of 60 public benchmarks tested by the developers. It excels in agent-centric tasks, including graphical user interface control and gameplay scenarios. In these areas, it reportedly outperforms systems such as OpenAI CUA and Claude 3.7. The model also demonstrates robust reasoning skills, particularly in solving visual puzzles. It handles visual and video understanding alongside complex multimodal reasoning challenges.

How to use it

Users can access Seed1.5-VL through the Volcano Engine platform. The specific model identifier provided is doubao-1-5-thinking-vision-pro-250428. The developers have made the technical report publicly available on arXiv to facilitate further research and inspection of their methods. No open weights or local deployment options are mentioned in the provided documentation.

Questions people ask about Seed1.5-VL

What type of model is Seed1.5-VL?

It is a vision-language foundation model that processes text, images, and video. It uses a Mixture-of-Experts architecture for its language component, allowing it to handle multimodal understanding and reasoning tasks efficiently.

How many parameters does the model use?

The language model component has 20 billion active parameters due to its Mixture-of-Experts design. The separate vision encoder contains 532 million parameters. This combination aims to balance performance with computational efficiency.

How does it perform against other models?

Developers claim it achieves state-of-the-art results on 38 out of 60 public benchmarks. It specifically outperforms leading systems like OpenAI CUA and Claude 3.7 in agent-centric tasks such as GUI control and gameplay.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Seed1.5-VL yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2505.07062. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.