ByteDance-Seed/Seed-OSS-36B-Instruct

Seed-OSS-36B-Instruct comes from ByteDance's Seed Team as part of the broader Seed-OSS family, a set of open-source large language models released under the Apache 2.0 license. The card describes the lineup as built for long-context handling, reasoning, agent workflows, and general-purpose use, and notes that the series was trained with 12T tokens while still posting strong results across popular open benchmarks. ByteDance also flags that Seed-OSS is primarily optimized for international multilingual scenarios, with a project page and a forthcoming technical report listed as additional resources for deeper context.

What sets the Instruct variant apart in practice is a controllable thinking budget, letting developers tune how much internal reasoning the model spends before producing an answer. Default thinking is unlimited, but setting a budget at multiples of 512 tokens (such as 512, 1K, 2K, 4K, 8K, or 16K) is recommended because those intervals were used heavily during training, and any budget below 512 is best collapsed to zero to force a direct response. On the deployment side, it is a 36B-parameter instruction-tuned model that is commonly served through an OpenAI-compatible API on vLLM-optimized H100 infrastructure, which makes it approachable for teams that already use OpenAI-style clients and want long-context reasoning with adjustable compute budgets.