ByteDance open-sources Seed-OSS-36B

ByteDance’s Seed team released Seed-OSS-36B, a 36-billion-parameter open-weight language model, under the permissive Apache 2.0 licence. The model natively supported context windows up to 512,000 tokens without the extension tricks many models rely on, and was trained on a comparatively modest 12 trillion tokens for its scale. ByteDance released both an instruction-tuned version and a base version without added synthetic instruction data, aimed at researchers who wanted to build their own post-training on top.

The release’s distinguishing feature was a “thinking budget”: users could set how many tokens the model was allowed to spend reasoning before answering, letting the same model trade off latency and cost against accuracy on a given task rather than being fixed at one reasoning depth. ByteDance reported strong results for the model’s size on knowledge, maths and coding benchmarks, and pitched it specifically at agentic and tool-use tasks.

Seed-OSS-36B mattered less for its scores than for what ByteDance chose to open. The company’s flagship conversational model, Seed 1.6, remained closed and API-only; Seed-OSS was released alongside it as ByteDance’s first substantial open-weight contribution, joining a wave of Chinese labs — Alibaba’s Qwen, DeepSeek, Moonshot’s Kimi and others — publishing competitive open-weight models through 2025 even as their most capable systems stayed proprietary. The split mirrored a pattern common across the industry: open releases used to build developer mindshare and research goodwill, closed releases used to monetise the frontier.