Seed-OSS-36B

2025-08-20 · Publisher : ByteDance · Domain : Artificial Intelligence

ByteDance releases Seed-OSS-36B, its first large general-purpose open-weights model: a dense model with 36 billion parameters, a native 512,000-token context and an adjustable thinking budget, under Apache 2.0.

### Context In August 2025, the Chinese open ecosystem is dominated by MoE models (Qwen3, DeepSeek, Kimi K2) and OpenAI has just opened gpt-oss. ByteDance, until then known for specialized open models (UI-TARS interface agents, Seed-Coder code model, BAGEL image model), publishes a general-purpose model with its Seed team. It is dense rather than MoE and optimized primarily for international use, according to the model card. ### What the model brings - Adjustable thinking budget: the user sets the length of the reasoning; the model self-assesses periodically (for example, with a 512-token budget it reports the tokens consumed and remaining) and concludes when the budget is exhausted. - Native 512,000-token context: score of 94.6 on RULER at 128,000 tokens, on par with Seed1.6-Thinking-0715 and Qwen3-30B-A3B-Thinking-2507 (94.5). - Data efficiency: trained on only 12 trillion tokens, according to the model card. - Research base: two base versions, with and without synthetic instruction data in pretraining, so that post-training research is not biased. - Agents: 56 on SWE-Bench Verified (OpenHands), versus 31 for Qwen3-30B-A3B-Thinking and 23.4 for Qwen3-32B, and 70.4 on TAU1-Retail. ### Architecture and specifications - Architecture: dense model with 36 billion parameters, 64 layers, grouped-query attention (80 query heads, 8 key and value heads), head size 128, hidden size 5,120, RoPE, RMSNorm and SwiGLU. - Vocabulary: 155,000 tokens. Context: 512,000 tokens. Training: 12 trillion tokens. Modalities: text. - License: Apache 2.0 (free commercial use). - Weights: Seed-OSS-36B-Instruct, Seed-OSS-36B-Base and Seed-OSS-36B-Base-woSyn on Hugging Face; compatible with vLLM. Knowledge cutoff: not disclosed. ### Cost and access No dedicated API or pricing was consulted for Seed-OSS. For self-hosting, the model card provides a vLLM command with tensor parallelism 8 and offers 4-bit and 8-bit quantization options; it does not set a minimum configuration. | Item | Information | | --- | --- | | Official API | Not consulted | | Self-hosting | Dense model with 36 billion parameters; vLLM example with tp 8; 4-bit and 8-bit quantization possible | | License | Apache 2.0 | ### Benchmarks | Benchmark | Seed-OSS-36B-Instruct | Seed1.6-Thinking-0715 | OAI-OSS-20B (open) | Qwen3-30B-A3B-Thinking-2507 (open) | | --- | --- | --- | --- | --- | | MMLU-Pro | 82.7 | 86.6 | 76.2 | 81.9 | | GPQA-Diamond | 71.4 | 80.7 | 72.2 | 71.4 | | AIME 2024 | 91.7 | 90.3 | 92.7 | 87.7 | | AIME 2025 | 84.7 | 86.0 | 90.3 | 81.3 | | LiveCodeBench v6 | 67.4 | 66.8 | 63.8 | 60.3 | | SWE-Bench Verified (OpenHands) | 56.0 | 41.8 | 60.7 | 31.0 | | TAU1-Retail | 70.4 | 63.0 | 54.8 | 58.7 | | RULER (128K) | 94.6 | 94.5 | 78.7 | 94.5 | Source: Hugging Face model card of Seed-OSS-36B-Instruct. Scores published by the Seed team; figures from other vendors are taken from public reports according to the model card. First figure retained when the model card gives two. ### Versus closed models In the model card's table, the only closed model is ByteDance's own Seed1.6-Thinking-0715: Seed-OSS-36B-Instruct outperforms it on AIME 2024 (91.7 vs 90.3), LiveCodeBench v6 (67.4 vs 66.8), SWE-Bench Verified (56.0 vs 41.8) and TAU1-Retail (70.4 vs 63.0), but remains behind on knowledge (MMLU-Pro 82.7 vs 86.6; GPQA-Diamond 71.4 vs 80.7). No closed model from OpenAI, Anthropic or Google is in the comparison. Against the open gpt-oss-20b, it is ahead on RULER and TAU1-Retail but behind on SWE-Bench Verified and AIME 2025. The price gap cannot be quantified. For enterprises, the appeal is a mid-size dense model under Apache 2.0, with a thinking budget that makes it possible to cap the generation cost. ### Key takeaways Seed-OSS-36B is an open general-purpose model from ByteDance that favors reasoning control and long context over size. It is not compared with the large American closed models in the model card, which limits the assessment of its relative level.

Tags : LLM, Open Source, ByteDance, Seed, Apache 2.0, Contexte 512K, Budget de réflexion