Bytedance releases new open source seed -OS -36B model

In release of both synthetic and non-scientific versions of the CID-OS-36B-base model, the seed team demanded to balance practical performance with research flexibility.

Synthetic-Detta version, Trained with additional instruction data, continuously Provides strong score on standard benchmark And means a high-performance general-purpose option.

Non-Scientic Model, Conversely, leaves these growth, causing A Cleaner Foundation Synthetic instruction introduced by data.

By providing both, the team provides access to better results to the applicable users, while the researchers maintain a neutral base to study post -training methods.

meanwhile, Seed-OS-36 B-Intection Model It is different that it is Trained with instruction data Instead of fully served as a foundation model, to give priority to performance and instructions.

All three models are issued under Apache-2.2.0 license, allowing free use, modification and redistribution by researchers and developers working for enterprises.

That means They can be used to make an internal, internal/customer-support for a company or exterior/customer-support, without paying any licensing fee or for use of application programming interface (API).

let’s Going On Shipping of Summer 2025 Chinese Companies Powerful Open Source Model Shipping Openai was trying to catch its own open source with GPT -SS doubles released earlier this month.

Seed team status Seed-OS for International ApplicationsArgument, agent-such as performing functioning and versatility in multilingual settings.

The seed team formed in 2023 has created a foundation model that can serve both research and applicable use cases.

Design and core features

The architectural familiar design of the Seed -OS -36B connects the familiar design options such as the cause language modeling, the groved query meditation, the swiglu activation, the RMSNORM and the rope position encoding.

Each model carry 36 billion parameters in 64 layers and supports the terminology of 155,000 tokens.

One of the defined features is its Native long-contemporary capacity with a maximum length of 512,000 tokens, Designed to process extended documents and logic chains without performance loss.

This is double the length of Openai’s new GPT-5 model family and is Equal to the text of about 1,600 pages, Length of a Christian Bible.

Another different element is the introduction of one Thinking budgetWhich allows developers to specify how much the model should argue before answering.

This is something that we have seen from other recent open source models, including Nvidia’s new Namotron-9 B-V2 Available for hugging,

In practice, this means that teams can tune the performance based on the complexity of the work and the efficiency requirements of deployment.

The budget is recommended in the multiples of 512 tokens, providing a direct response mode with 0/

Competitive performance on third-party benchmark

The benchmark release positions were published among the strong large open-source models with the position Seed-OS-36B. The instruction version, in particular, posts state -of -the -art results in many areas.

• Mathematics and logic: Seed-OS-36 B-Astract 91.7 percent on Aime24 And 65 on beorimeBoth open-sources represents “state-of-the-art”.
• Coding: On Livecodebench V6, instruction model record 67.4Another Sota Score.
• Prolong: 128K reference to the ruler at length, it reaches 94.6The highest open-source result reported.
• Base model performance: Synthetic-Deta delivered the base variant 65.1 mmlu-pro And 81.7 on MathematicsBoth state -of -the -arts result in results in their categories.

The no-science base version, while slightly back over several measures, proves to be competitive in itself.

it GPQA-D improves its synthetic counterpart, Providing researchers with a cleaner, instruction-free base line for use.

For enterprises comparing open options, these results Seed-OS suggests that mathematics-thorough, coding and long-working workloads provide strong potential While still providing flexibility for cases of research use.

Accession and periphery

Beyond the performance, the seed team highlights access to developers and doctors. Model Hugging can be deployed using face transformerwith Both 4-bit and 8-bit formats To reduce memory requirements.

They too Integrate with VLLM for scalable servingConfiguration examples and API server instructions.

To reduce further obstacles, the team includes estimates, quick adaptation and script for equipment integration.

For Technical leaders who manage small teams are working under lack of budgetThese provisions are more acceptable to experiment with 36 billion-parameter models.

License and views for enterprise decision making

With models introduced under Apache-2.0, organizations can adopt them without restrictive licensing conditions, an important factor for teams that balance legal and operating concerns.

For decision makers evaluating open-source landscape, release brings three takeaways:

• Mathematics, coding, and long -term reference arguments across the state -of -the -art benchmark.
• A balance between high-performing synthetic-informed models and clean research base lines.
• Accessibility features that reduce operating overheads for lean engineering teams.

By placing strong performances and flexible sinsitans under an open license, the seed team of bidence adds new options for enterprises, researchers and developers.

Daily insights on business use cases with VB daily

If you want to impress your boss, VB daily has covered you. We give you the scoop inside what companies are doing with generative AI, from regulatory changes to practical deployment, so you can share insight for maximum ROI.

Read our privacy policy

Thanks for membership. See more VB newsletters here.

There was an error.