In releasing each artificial and non-synthetic variations of the Seed-OSS-36B-Base mannequin, the Seed Group sought to stability sensible efficiency with analysis flexibility.

The synthetic-data variant, skilled with extra instruction information, persistently delivers stronger scores on normal benchmarks and is meant as a higher-performing general-purpose choice.

The non-synthetic mannequin, against this, omits these augmentations, creating a cleaner basis that avoids potential bias or distortion launched by artificial instruction information.

By offering each, the group provides utilized customers entry to improved outcomes whereas guaranteeing researchers retain a impartial baseline for learning post-training strategies.

In the meantime, the Seed-OSS-36B-Instruct mannequin differs in that it’s post-trained with instruction information to prioritize activity execution and instruction following, quite than serving purely as a basis mannequin.

All three fashions are launched beneath the Apache-2.0 license, permitting free use, modification, and redistribution by researchers and builders working for enterprises.

Which means they can be utilized to energy industrial functions, inside to an organization or exterior/customer-facing, with out paying ByteDance any licensing charges or for utility programming interface (API) utilization.

This continues the summer 2025 trend of Chinese companies shipping powerful open source models with OpenAI making an attempt to meet up with its own open source gpt-oss duet released earlier this month.

The Seed Group positions Seed-OSS for worldwide functions, emphasizing versatility throughout reasoning, agent-like activity execution, and multilingual settings.

The Seed Group, shaped in 2023, has targeting constructing basis fashions that may serve each analysis and utilized use circumstances.

Design and core options

The structure behind Seed-OSS-36B combines acquainted design selections similar to causal language modeling, grouped question consideration, SwiGLU activation, RMSNorm, and RoPE positional encoding.

Every mannequin carries 36 billion parameters throughout 64 layers and helps a vocabulary of 155,000 tokens.

One of many defining options is its native long-context functionality, with a most size of 512,000 tokens, designed to course of prolonged paperwork and reasoning chains with out efficiency loss.

That’s twice the size of OpenAI’s new GPT-5 model family and is roughly equal to about 1,600 pages of textual content, the size of a Christian Bible.

One other distinguishing component is the introduction of a pondering funds, which lets builders specify how a lot reasoning the mannequin ought to carry out earlier than delivering a solution.

It’s one thing we’ve seen from different latest open supply fashions as nicely, together with Nvidia’s new Nemotron-Nano-9B-v2, additionally available on Hugging Face.

In follow, this implies groups can tune efficiency relying on the complexity of the duty and the effectivity necessities of deployment.

Budgets are really helpful in multiples of 512 tokens, with 0 offering a direct response mode/

Aggressive efficiency on third-party benchmarks

Benchmarks printed with the discharge place Seed-OSS-36B among the many stronger massive open-source fashions. The Instruct variant, particularly, posts state-of-the-art ends in a number of areas.

• Math and reasoning: Seed-OSS-36B-Instruct achieves 91.7 % on AIME24 and 65 on BeyondAIME, each representing open-source “state-of-the-art” (SOTA).
• Coding: On LiveCodeBench v6, the Instruct mannequin information 67.4, one other SOTA rating.
• Lengthy-context dealing with: On RULER at 128K context size, it reaches 94.6, marking the best open-source outcome reported.
• Base mannequin efficiency: The synthetic-data Base variant delivers 65.1 on MMLU-Professional and 81.7 on MATH, each state-of-the-art ends in their classes.

The no-synthetic Base model, whereas barely behind on many measures, proves aggressive in its personal proper.

It outperforms its artificial counterpart on GPQA-D, offering researchers with a cleaner, instruction-free baseline for experimentation.

For enterprises evaluating open choices, these outcomes recommend Seed-OSS affords robust potential throughout math-heavy, coding, and long-context workloads whereas nonetheless offering flexibility for analysis use circumstances.

Entry and deployment

Past efficiency, the Seed Group highlights accessibility for builders and practitioners. The fashions may be deployed utilizing Hugging Face Transformers, with quantization assist in each 4-bit and 8-bit codecs to scale back reminiscence necessities.

In addition they combine with vLLM for scalable serving, together with configuration examples and API server directions.

To decrease limitations additional, the group contains scripts for inference, immediate customization, and gear integration.

For technical leaders managing small groups or working beneath funds constraints, these provisions are positioned to make experimentation with 36-billion-parameter fashions extra approachable.

Licensing and concerns for enterprise decision-makers

With the fashions supplied beneath Apache-2.0, organizations can undertake them with out restrictive licensing phrases, an necessary issue for groups balancing authorized and operational considerations.

For resolution makers evaluating the open-source panorama, the discharge brings three takeaways:

• State-of-the-art benchmarks throughout math, coding, and long-context reasoning.
• A stability between higher-performing synthetic-trained fashions and clear analysis baselines.
• Accessibility options that decrease operational overhead for lean engineering groups.

By putting robust efficiency and versatile deployment beneath an open license, ByteDance’s Seed Group has added new choices for enterprises, researchers, and builders alike.