TikTok has garnered attention recently as the White House joined its platform, while its parent company, ByteDance, announced the release of a new AI model. The Seed Team, comprising AI researchers from ByteDance, unveiled Seed-OSS-36B on the AI code-sharing site Hugging Face.
Seed-OSS-36B is a collection of open-source large language models (LLMs) designed for advanced reasoning and developer usability. It offers a longer token context length compared to many U.S.-based competitors, including OpenAI and Anthropic. The model collection consists of three primary variants: Seed-OSS-36B-Base (with and without synthetic data) and Seed-OSS-36B-Instruct, which focuses on task execution.
The Base model comes in two versions. The synthetic-data version, which includes additional instruction data, aims for higher performance, while the non-synthetic variant provides a neutral foundation, allowing researchers to study outcomes without biases. The Instruct model is optimized for instruction following.
All three models are available under the Apache-2.0 license, allowing free use, modification, and redistribution for commercial applications without incurring licensing fees. This aligns with a broader trend of Chinese firms releasing open-source models, as companies like OpenAI have also begun to explore open-source options.
The Seed Team has tailored Seed-OSS to support varied applications, with prominent features such as a maximum context capability of 512,000 tokens, enabling the processing of extensive documents. The architecture includes a “thinking budget,” allowing developers to control the model’s reasoning depth per task.
Benchmark results indicate the Seed-OSS-36B models exhibit competitive performance in math, coding, and long-context reasoning. The synthetic Base variant scored strong results on benchmarks, while the non-synthetic version also achieved notable performance.
Seed-OSS models prioritize accessibility for developers, offering features like quantization support and integration with existing tools. Their open licensing terms facilitate easier deployment for organizations while balancing high performance with operational flexibility.
Source: https://venturebeat.com/ai/tiktok-parent-company-bytedance-releases-new-open-source-seed-oss-36b-model-with-512k-token-context/