News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnSeed-OSS-36B

Seed-OSS-36B

Open-source 36B parameter language model with controllable reasoning.

ByteDance's Seed Team released Seed-OSS-36B in August 2025 as an open-source large language model. It features 36 billion parameters and supports a native context window of 512,000 tokens. The model allows users to control the length of its reasoning process, balancing speed and accuracy for complex tasks.

Share on WhatsApp

  • Released by ByteDance Seed Team in August 2025.
  • Supports a native context window of 512,000 tokens.
  • Users can set a specific token budget for reasoning.
  • Licensed under Apache-2.0 for open use.
  • Trained on 12 trillion tokens of data.

What it is

Seed-OSS-36B is a series of open-source large language models created by the ByteDance Seed Team. Released in August 2025, the family includes base and instruction-tuned variants with 36 billion parameters. It is designed for general purposes, with specific optimizations for long-context understanding, logical reasoning, and agentic tasks such as using external tools. The model is trained on 12 trillion tokens and supports a native context window of up to 512,000 tokens.

How it came about

The ByteDance Seed Team, founded in 2023, developed the Seed-OSS series to contribute to the open-source community. On 20 August 2025, the team released Seed-OSS-36B-Base and Seed-OSS-36B-Instruct. To support research into the effects of synthetic data on pre-training, they also provided a base model variant trained without synthetic instruction data, known as Seed-OSS-36B-Base-woSyn.

Inside the model

The model uses a causal decoder architecture featuring Rotary Positional Embeddings (RoPE), Grouped Query Attention (GQA), RMSNorm, and SwiGLU activation functions. It has 80 query heads and 8 key-value heads. A distinctive feature is the flexible thinking budget, which allows users to limit the number of tokens the model uses for internal reasoning. This helps manage inference efficiency. The model performs self-reflection during this process to track its remaining token allowance.

What it can do

Seed-OSS-36B is optimized for international use cases and demonstrates strong performance in reasoning and agent-based tasks. It handles tool usage and issue resolution effectively. Benchmarks indicate competitive results on coding tests like LiveCodeBench and agent evaluations such as SWE-Bench. The model maintains balanced general capabilities while excelling in tasks that require extended chains of thought.

How to use it

The weights are available under the Apache-2.0 licence, permitting broad commercial and research use. Developers can download the checkpoints from Hugging Face. Inference is supported via the Transformers library and vLLM, with options for 4-bit and 8-bit quantization to reduce memory requirements. The team recommends specific sampling parameters, such as a temperature of 1.1, for optimal generation.

Questions people ask about Seed-OSS-36B

What is the thinking budget feature?

It allows users to limit the number of tokens the model uses for internal reasoning. This helps control inference time and cost. If no budget is set, the model reasons without limit. Setting it to zero forces a direct response.

Is Seed-OSS-36B free to use?

Yes, the model weights are released under the Apache-2.0 licence. This permissive licence allows for both commercial and non-commercial use, modification, and distribution without significant restrictions.

How large is the context window?

The model supports a native context window of up to 512,000 tokens. This allows it to process and reason over very long documents or codebases without losing track of earlier information.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Seed-OSS-36B yet, so the facts on this page come from the sources listed above: huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.