What is Seed1.6
Seed1.6 is a general model series launched by the ByteDance Seed team, integrating multimodal capabilities and supporting deep reasoning with a 256K long context. Seed1.6 continues the sparse MoE exploration results of Seed1.5, and through three stages—pure text pretraining, multimodal mixed continual training, and long-context continual training—improves text and visual capabilities. In the post-training stage, reasoning capabilities are strengthened, and Seed1.6-Thinking and Seed1.6 (Adaptive CoT) are developed to achieve a balance between extreme reasoning and dynamic thinking. It performs outstandingly in generalization tests such as the Gaokao and JEE Advanced. In the future, it will explore more efficient architectures, improve reasoning effectiveness, and enrich multimodal capabilities.
Main Features of Seed1.6
- Multimodal Understanding: It can simultaneously process text and visual information, understand and analyze visual content such as images and videos, and enable interactions that combine text and images.
- Deep Reasoning: It supports deep reasoning with a 256K long context and can handle complex logic problems and long-text tasks, such as long-form reading comprehension and multi-step reasoning.
- Adaptive Thinking: It automatically selects the thinking mode according to question difficulty (full thinking, no thinking, adaptive thinking), balancing reasoning effectiveness and performance.
- Graphical Interface Operation: It supports understanding and operating graphical interfaces, such as web pages and software interfaces, to enable automated tasks and interactive operations.
Technical Principles of Seed1.6
- Pretraining: It uses data such as web pages, books, papers, and code for training to improve the quality and knowledge density of pretraining data. Based on increasing the proportion of subject, code, and reasoning data, visual modality data is added for mixed training with high-quality text data. Using long-text data of different length ranges, the model's maximum sequence length is gradually increased from 32K to 256K.
- Post-training：
- Seed1.6-Thinking: Based on longer thinking processes, it achieves extreme reasoning effectiveness and uses multi-stage RFT and RL iterative optimization to increase the model's thinking length on complex problems, deeply integrating VLM to bring clear visual understanding capabilities.
- Seed1.6（Adaptive CoT）: Based on dynamic thinking technology, it compresses CoT length while ensuring effectiveness, achieving a dynamic balance between performance and effectiveness. A new reward function is introduced, allowing the model to automatically choose whether to think based on different prompts.
- Architecture and Algorithm Improvements: It continuously improves the model architecture, training algorithms, and infrastructure to enhance model performance and efficiency. It uses efficient data cleaning, filtering, deduplication, and sampling strategies to improve the quality of pretraining data. Based on parallel decoding technology, it uses more thinking tokens before giving an answer to improve the model's performance on high-difficulty tasks.
Performance of Seed1.6
- MMLU Test: In the MMLU test, Seed1.6-AdaCoT has a CoT trigger rate of 37%, effectively saving tokens without performance degradation.
- Gaokao Test: In the 2025 Shandong paper Gaokao real-question test, Seed1.6-Thinking scored 683 in the liberal arts and 648 in the sciences, exceeding the previous years' admission score lines of most 985 universities.
- AIME Test: In the AIME test, Seed1.6-AdaCoT has a CoT trigger rate of 90%, with effectiveness comparable to Seed1.6-FullCoT.
- BeyondAIME Test: In the BeyondAIME test, Seed1.6-Thinking achieved an 8-point improvement in test results based on parallel decoding.
- JEE Advanced Test: Seed1.6-Thinking achieved a top 10 all-India score in the JEE Advanced exam test, answering all questions correctly in the mathematics test.
Seed1.6 project address
- Project official website：https://exp.volcengine.com/ark?model=doubao-seed-1-6-250615
Seed1.6 application scenarios
- Education: Provide students with personalized learning tutoring, automatically adjust the depth of thinking, give detailed solution steps and feedback, process exam answers, and score accurately.
- Content creation: Generate high-quality copy, support long-form text creation, assist designers with creative inspiration and optimization suggestions, and improve content creation efficiency.
- Smart office: Automatically analyze long documents, extract key information to generate summaries, serve as an intelligent assistant to handle daily office tasks, and improve office efficiency.
- Healthcare: Combine text and imaging to assist doctors in diagnosis, provide analysis reports, read medical literature to extract key information, and support medical research and clinical decision-making.
- Intelligent customer service: Automatically adjust the depth of thinking according to question complexity, provide solutions, analyze user emotions, and improve the customer service experience.