Seed-OSS

Seed-OSS is ByteDance Seed Team's open-source large language model series, released August 20, 2025. It ships in three variants: Seed-OSS-36B-Base trained with synthetic instruction data, Seed-OSS-36B-Base-woSyn as a clean baseline without it, and Seed-OSS-36B-Instruct post-trained for instruction following. At 36.2 billion parameters with a 512K-token context window and a controllable thinking budget, it was trained on only 12 trillion tokens.

The Instruct variant reports strong open-source results in reasoning, math, and coding, including 81.7 on MATH, and the whole series is released under Apache 2.0, so it can power commercial applications without licensing fees. ByteDance positions it for international use cases with strong multilingual support.

The honest caveats: benchmark numbers are vendor-reported, and the two Base variants serve different audiences, with the synthetic-data version scoring higher but the clean version preferred for research on post-training. Weights are on Hugging Face and the release repo is on GitHub.