ByteDance Releases Seed-OSS-36B With Open Licensing and a 512K Context Window

A 36 Billion Parameter Model Enters the AI Contest

ByteDance, the Chinese technology company behind TikTok and Douyin, has released Seed-OSS-36B, an artificial intelligence model with 36 billion parameters, a context window of 512,000 tokens and an Apache-2.0 license that permits commercial use. As of August 22, 2025, the release places the company in closer competition with DeepSeek, Alibaba Cloud, Google and OpenAI.

The model comes from ByteDance's Seed Team, which develops foundation models, the underlying AI systems that developers can adapt for many different tasks. Seed-OSS-36B is a family of large language models, or LLMs, designed to understand and generate text. Its uses extend from answering questions to coding, research and analysis of lengthy documents.

Three features define the release: the ability to work with unusually large amounts of text at once, a configurable thinking budget that controls how much reasoning the model performs before answering, and permission for developers to download, modify and build commercial products around it.

ByteDance is also presenting the model as an example of efficient training. Seed-OSS-36B was trained with 12 trillion tokens, less than many competitors, while recording strong results on tests of knowledge, mathematics and programming. Its parameter count is smaller than that of some rival systems.

For businesses and research teams, the practical attraction is control. They can adapt the model to their needs and choose how to deploy it, rather than relying entirely on access to a proprietary service. The release also adds another competitor to an increasingly crowded market in which Chinese technology companies are using open models to attract developers.

What the Model's Size and License Mean

Seed-OSS-36B's 36 billion parameters are internal variables that help it learn patterns and generate text. The number describes the scale of the model, while its usefulness also depends on its architecture, training and ability to follow instructions.

The Apache-2.0 license allows developers and businesses worldwide to use, modify and build on Seed-OSS-36B for free, including for commercial purposes. Smaller organizations and research groups can therefore work with the model without paying a fee for permission to use it.

Proprietary models often restrict access or charge usage fees. ByteDance's approach gives users greater freedom to adapt the underlying system, making the release relevant to companies that want more control over their AI applications.

The company has made the Seed-OSS-36B-Instruct model available on Hugging Face, described as the world's largest platform for open source AI. Distribution through the platform supports experimentation, collaboration and scrutiny by developers around the world.

How Much Can a 512K Context Window Hold?

Seed-OSS-36B can process up to 512,000 tokens in a single session, an amount described as roughly 1,600 pages of text. Tokens are the units of text that a language model processes. Its context window is the amount it can work with at once when producing an answer.

The 512K capacity is four times the 128K context window of models such as DeepSeek V3.1. That comparison helps explain the appeal: a larger window allows a user to supply much more material in one request, rather than divide it into smaller pieces.

Legal teams could submit entire case files for analysis. Healthcare researchers could work through years of patient records together. Software developers could examine or refactor whole codebases, while academics could review lengthy reports or datasets without repeatedly separating the material.

ByteDance describes the long context capability as native to the model. Seed-OSS-36B supports it without relying on later changes or interpolation to extend its reach. In practical terms, the ability to handle long inputs is part of the model's technical design.

The feature broadens the kinds of assignments developers can attempt. An AI application working on a large collection of related documents can receive more of that material together, giving it a wider body of information to draw on during the task.

A Thinking Budget Trades Speed for More Reasoning

Seed-OSS-36B gives developers a thinking budget, a setting that limits how much reasoning the model performs before delivering a response. The budget is measured in tokens and can be set in multiples of 512.

A small budget is intended to produce a faster answer for a straightforward request. A larger budget allows more reasoning for a difficult problem, with the aim of producing a more accurate response. The setting lets developers adjust the balance between speed and depth for different jobs.

Customer support applications, for example, may need quick responses to routine questions. Financial analysis may call for more extended reasoning. The same model can be configured differently for those tasks, rather than treating every request as equally demanding.

Seed-OSS-36B also targets agentic intelligence. Here, that means the ability to use tools, work through several steps and solve problems as part of a larger task. Research, coding and legal analysis are among the applications identified for those capabilities.

The combination of a large context window and a configurable reasoning budget gives developers two separate controls: how much information the model receives and how much reasoning it performs before answering. Those choices are central to adapting it for different workloads.

What the Benchmark Figures Show

Seed-OSS-36B recorded a score of 65.1 on MMLU-Pro, a test of knowledge and reasoning, compared with 58.5 for Alibaba's Qwen2.5-32B-Base. The difference is 6.6 points. The model also achieved high marks on TriviaQA and BBH reasoning tests.

The mathematics and coding results cover several different tasks. Seed-OSS-36B scored 90.8 on GSM8K and 81.7 on MATH. On programming tests, it scored 76.8 on HumanEval and 80.6 on MBPP.

• MMLU-Pro: 65.1, compared with 58.5 for Qwen2.5-32B-Base.
• GSM8K: 90.8.
• MATH: 81.7.
• HumanEval: 76.8.
• MBPP: 80.6.
• AIME24: 91.7 for Seed-OSS-36B-Instruct.

The instruction tuned version, Seed-OSS-36B-Instruct, scored 91.7 on the AIME24 mathematics competition benchmark, placing second only to OpenAI's OSS-20B in the comparison described at release. Instruction tuning prepares a model to respond to requests rather than simply continue a passage of text.

The results give ByteDance a basis for competing with models of similar size on knowledge, reasoning and technical work. The comparison with Alibaba is especially concrete: both the model name and the MMLU-Pro scores are given, allowing readers to see the size of the reported gap.

Google and OpenAI also form part of the wider competitive field. Seed-OSS-36B's pitch combines its test results with open licensing, long context processing and adjustable reasoning, rather than relying on parameter count alone.

Inside the Architecture and Training

Seed-OSS-36B uses a dense transformer architecture with 64 layers and a hidden size of 5,120. Layers are successive stages of processing, while hidden size describes the width of the model's internal representations. These are part of the structure through which it processes text.

The architecture includes RoPE, or Rotary Position Embedding, for representing the position of information in a sequence. It also uses GQA, or Grouped Query Attention, to make its attention mechanisms more efficient. Attention is the process that helps a language model connect relevant parts of its input.

RMSNorm provides normalization, and SwiGLU supplies the activation function used in the model. These components help organize how information passes through the network. Seed-OSS-36B's vocabulary contains 155,000 tokens, supporting a wide range of languages and tasks.

The model was trained on 12 trillion tokens. Its strong performance with less training data than many competitors is part of ByteDance's efficiency argument: training methods and architecture can help a model compete even when it has fewer parameters or a smaller training volume.

ByteDance provides two base versions. One includes synthetic instruction data for stronger performance; the other leaves that data out to support purer research. The choice gives applied developers and academic researchers different starting points according to what they want to build or investigate.

What Developers Receive for Deployment

ByteDance provides quantized versions, deployment scripts and compatibility with popular AI frameworks. Together, those resources are intended to make Seed-OSS-36B easier to test and put into service.

Quantized versions are smaller, faster forms of the model. Deployment scripts help users set it up, while framework compatibility allows it to work with tools developers already use. These resources matter because a model's practical value depends partly on how readily a team can integrate it into an application.

Researchers can choose between the base versions, while developers seeking a model prepared to follow requests can use the instruction tuned variant. The release therefore supports different kinds of work, from studying model behavior to building products around it.

The flexibility also connects the technical package to the license. Permission to modify the model gives teams room to adapt it; deployment resources help them begin that work. For smaller organizations, that combination lowers the barrier to experimenting with an advanced language model.

Why ByteDance Is Giving Developers Access

ByteDance's decision to release Seed-OSS-36B openly is a business strategy as well as a technical release. Broad access can encourage developers, researchers and companies to build applications around its technology.

More adoption could increase ByteDance's influence across the AI ecosystem. Feedback from users can help the company improve its models, and successful innovations developed around the release could potentially find their way into its own products.

Open licensing also puts competitive pressure on providers that rely on paid or restricted access. Western companies such as OpenAI and Google often charge for API access or restrict commercial use. An API is a connection through which an application sends requests to an AI service and receives responses.

A benchmark comparison report describes businesses using open source AI tools as often reporting higher returns on investment than those using proprietary models, citing lower costs and greater flexibility. That business argument sits alongside the technical case for Seed-OSS-36B.

Open releases are also part of the broader effort by Chinese AI companies to close the performance gap with the United States. Making models available to a wider developer community can speed adoption and allow more people to contribute improvements.

Chinese Rivals Accelerate Their Own Releases

Alibaba's shares traded lower following the Seed-OSS-36B announcement, reflecting investor concerns about stronger competition in AI. The release arrived during a broader period of rapid development among Chinese technology companies.

DeepSeek responded by upgrading its flagship V3 model, with an emphasis on faster processing and optimization for chips made in China. The focus on domestic semiconductors is important amid continuing US export controls on advanced chips.

Alibaba, Baidu, Tencent and Huawei are also racing to develop more powerful and efficient AI systems. Open source strategies are central to that effort because they can bring models into more products and draw on improvements from a wider community.

A Chinese AI market report describes Chinese technology stocks rallying on AI advances, with gains tempered by global economic conditions and regulatory challenges. Alibaba's decline following the announcement sits within that broader market picture.

The competition therefore runs on several tracks at once: benchmark performance, access to suitable chips, ease of deployment and the ability to attract developers. Seed-OSS-36B strengthens ByteDance's presence across that contest with a model developers can obtain and adapt.

Early Reactions Focus on Control and Flexibility

Industry analysts and AI practitioners have praised Seed-OSS-36B's efficiency and performance. A LinkedIn model analysis described it as a "game changer," pointing to its long context window, open licensing and configurable thinking budget.

The analyst behind that assessment argued that the release gives organizations greater choice over how they use AI:

With longer context, open-source licensing, and thinking budgets, businesses gain more control and flexibility. ByteDance has provided a powerful tool for free, and it's now up to businesses, researchers, and developers to decide how to use it.

Early users discussing the model on Reddit have also highlighted its versatility and options suited to research. Those reactions focus on the ability to choose between versions, adjust reasoning and work with large amounts of material.

For a business evaluating the release, the concrete decisions are whether its tasks benefit from a 512K context window, how much reasoning each request needs and which version best suits its work. Researchers have the additional choice of a base model trained with or without synthetic instruction data.

What the Release Changes for AI Users

Seed-OSS-36B brings several trends together in one release. Its benchmark results support a focus on efficient architecture and training alongside model size. Its license expands access, while its context window and thinking budget create room for applications involving extensive documents and complex tasks.

Legal analysis, scientific research, software development and work with medical records are among the areas identified for possible use. Across those fields, the shared attraction is the ability to process more information together and adjust the model's reasoning to the job.

Developers gain another model they can modify and deploy. Researchers gain alternative base versions. Competing AI providers face another openly available system seeking adoption through performance, flexibility and control.

ByteDance's release adds momentum to an industry in which Chinese and US companies are competing on both capability and access. The immediate change is tangible: a model with 36 billion parameters, a 512K context window and commercial use permissions is available for teams to build on.

Key Points

• ByteDance has released Seed-OSS-36B, developed by its Seed Team, with 36 billion parameters.
• The 512,000 token context window is roughly equivalent to 1,600 pages and four times the 128K window of DeepSeek V3.1.
• Apache-2.0 licensing permits free use, modification and commercial development.
• A thinking budget, adjustable in multiples of 512 tokens, lets developers balance response speed and reasoning depth.
• The model scored 65.1 on MMLU-Pro, against 58.5 for Qwen2.5-32B-Base; the instruction tuned version scored 91.7 on AIME24.
• Training used 12 trillion tokens, and two base versions support different applied and research needs.
• Quantized versions, deployment scripts and framework compatibility support experimentation and deployment.
• The release increases competition involving Alibaba, DeepSeek, Google and OpenAI as Chinese companies expand their AI efforts.