Seed-OSS

Seed-OSS is ByteDance Seed's entry in the open-weight LLM race, led by a 36B-parameter model trained on only 12T tokens yet posting strong open benchmarks, released in August 2025 under Apache-2.0. The architecture is deliberately conventional, a causal transformer with RoPE, grouped-query attention, RMSNorm, and SwiGLU, but it was trained natively with up to 512K tokens of context. Its signature feature is a controllable thinking budget: users cap reasoning length in multiples of 512 tokens, and the model periodically self-reflects on remaining budget during chain-of-thought, giving explicit control over the latency-quality tradeoff. The release includes an Instruct model plus two base variants, Seed-OSS-36B-Base and Seed-OSS-36B-Base-woSyn, trained with and without synthetic instruction data, an unusual gift to post-training researchers who want an uncontaminated base. ByteDance highlights agentic strength on tool use and issue resolution. Weights are on Hugging Face with vLLM and Transformers support, serving at long context calls for a large-memory GPU setup, and Apache-2.0 leaves commercial use unrestricted.