Seed 1.6 - A general model series launched by ByteDance
Seed 1.6 is a general-purpose model series launched by ByteDance's Seed team. It integrates multimodal capabilities and supports deep inference with 256K long contexts. Seed 1.6 utilizes the sparse MoE exploration results of Seed 1.5, and undergoes pure text pre-training, multimodal hybrid...
What is Seed 1.6?
Seed 1.6 is a general-purpose model series launched by ByteDance's Seed team. It integrates multimodal capabilities and supports deep reasoning with 256K long contexts. Seed 1.6 inherits the sparse MoE exploration results of Seed 1.5, improving text and vision capabilities through three stages: pure text pre-training, multimodal hybrid continuous training, and long context continuous training. The post-training stage strengthens reasoning capabilities, developing Seed 1.6-Thinking and Seed 1.6 (Adaptive CoT) to achieve a balance between extreme reasoning and dynamic thinking. It has performed outstandingly in generalization tests such as the National College Entrance Examination (Gaokao) and JEE Advanced. Future development will explore more efficient architectures to improve reasoning performance and enrich multimodal capabilities.
Main functions of Seed 1.6
- Multimodal understandingIt can process text and visual information simultaneously, understand and analyze visual content such as images and videos, and achieve interactive displays that combine text and graphics.
- Deep reasoningIt supports deep reasoning with 256K long contexts and can handle complex logic problems and long text tasks, such as long reading comprehension and multi-step reasoning.
- Adaptive ThinkingAutomatically selects the thinking mode (full thinking, no thinking, adaptive thinking) based on the difficulty of the problem, balancing reasoning effectiveness and performance.
- Graphical interface operationIt supports understanding and operating graphical interfaces, such as web pages and software interfaces, to achieve automated tasks and interactive operations.
Technical Principles of Seed 1.6
- Pre-trainingThe model is trained using data from web pages, books, papers, and code to improve the quality and knowledge density of the pre-training data. By increasing the proportion of subject-specific, code, and reasoning-related data, visual modal data is mixed with high-quality text data for training. The maximum sequence length of the model is gradually increased from 32K to 256K using long text data of varying lengths.
- Post-training:
- Seed1.6-ThinkingAchieving optimal reasoning performance through a longer thought process, it uses multi-stage RFT and RL iterative optimization to enhance the model's thinking length on complex problems and deeply integrates VLM to bring clear visual understanding capabilities.
- Seed1.6 (Adaptive CoT)Based on dynamic thinking technology, the CoT length is compressed while maintaining effectiveness, achieving a dynamic balance between performance and effectiveness. A new reward function is introduced, allowing the model to automatically choose whether to engage in thinking based on different prompts.
- Architecture and algorithm improvementsContinuously improve model architecture, training algorithms, and infrastructure to enhance model performance and efficiency. Improve the quality of pre-training data using efficient data cleaning, filtering, deduplication, and sampling strategies. Enhance model performance on challenging tasks by using more thought tokens before providing the answer, based on parallel decoding technology.
Performance of Seed 1.6
- MMLU testSeed1.6-AdaCoT achieved a CoT trigger rate of 37% in the MMLU test, effectively saving tokens without performance degradation.
- College Entrance ExaminationSeed1.6-Thinking scored 683 points in the 2025 Shandong Gaokao (National College Entrance Examination) and 648 points in the science section, exceeding the admission scores of most 985 universities in previous years.
- AIME testingSeed1.6-AdaCoT achieved a CoT trigger rate of 90% in AIME testing, which is comparable to Seed1.6-FullCoT.
- BeyondAIME testSeed 1.6-Thinking improved the BeyondAIME test score by 8 points after implementing parallel decoding.
- JEE Advanced TestSeed1.6-Thinking achieved a top 10 score in India on the JEE Advanced test, answering all questions correctly on the math test.
Seed 1.6 project address
- Project official website: https://exp.volcengine.com/ark?model=doubao-seed-1-6-250615
Application scenarios of Seed 1.6
- EducationIt provides students with personalized learning guidance, automatically adjusts the depth of thinking, provides detailed problem-solving steps and feedback, processes exam answers, and accurately scores.
- Content creationGenerates high-quality copy, supports long text creation, assists designers with creative inspiration and optimization suggestions, and improves content creation efficiency.
- Smart OfficeIt automatically analyzes long documents, extracts key information to generate summaries, and acts as an intelligent assistant to handle daily office tasks, improving office efficiency.
- HealthcareIt combines text and images to assist doctors in diagnosis, provides analytical reports, extracts key information from medical literature, and supports medical research and clinical decision-making.
- Intelligent Customer ServiceIt automatically adjusts the depth of thinking based on the complexity of the problem, provides solutions, analyzes user emotions, and improves the customer service experience.