Particle.news
Download on the App Store
3 ARTICLES
·
13h ago

ByteDance’s Seed Team Open-Sources 36B-Parameter Seed-OSS Models With 512K Context

Open licensing removes commercial hurdles as performance figures await outside checks.

Image
Image
ByteDance’s Seed department oversees the social media giant’s artificial intelligence research and large language model development. Photo: Shutterstock
Image

Overview

  • ByteDance uploaded Seed-OSS-36B Base, Base-woSyn, and Instruct variants to Hugging Face under the Apache-2.0 license, enabling free commercial use.
  • The 36B-parameter design spans 64 layers with a 155,000-token vocabulary, native 512,000-token context, and a controllable thinking budget for reasoning length.
  • ByteDance-reported results claim open-source state-of-the-art for the Instruct model, including 91.7 on AIME24, 67.4 on LiveCodeBench v6, and 94.6 on RULER at 128K.
  • Documentation covers deployment via Transformers and vLLM, with 4-bit and 8-bit quantization options to reduce memory requirements.
  • The company says Seed-OSS-36B matches or surpasses similar-sized models from Alibaba Cloud, Google and OpenAI, with training reportedly using about 12 trillion tokens.