ByteDance Seed OSS 36B — open-source model for general-purpose chat and instruction following.
Seed-OSS-36B is an open-source 36B-parameter LLM released August 20, 2025 by ByteDance's Seed Team under the Apache-2.0 license. It is explicitly optimized for international (i18n) use cases and ships with three variants: Seed-OSS-36B-Base (w/ synthetic instruction data), Seed-OSS-36B-Base-woSyn (without synthetic data), and Seed-OSS-36B-Instruct. The model was trained on 12T tokens and supports a native 512K context window — double the length of many contemporary long-context models.
For TheRouter operators, Seed-OSS-36B is a strong open-weight reasoning and agentic model with a distinctive 'thinking budget' control: users can specify an integer multiple of 512 tokens (e.g., 512, 1K, 2K, 4K) to limit chain-of-thought length at inference time, improving efficiency on simpler tasks while preserving full reasoning depth when needed. It leads or matches open-source SOTA on AIME24/25, LiveCodeBench v6, SWE-Bench Verified, and RULER 128K, while remaining competitive on multilingual and general-knowledge benchmarks.
| Type | Rate |
|---|---|
| Input | $0.2268 / 1M tokens |
| Output | $0.6156 / 1M tokens |
| Release date | 2025-08-20github.com ↗ | verified |
| License | Apache 2.0 (weights & code)github.com ↗ | verified |
| Architecture | 36B parameters, 64 layers, GQA (80/8/8 QKV heads), RoPE, SwiGLU, RMSNorm, vocab 155Kgithub.com ↗ | verified |
| Context window | 512K tokens (native)github.com ↗ | verified |
| Training data | 12T tokensgithub.com ↗ | verified |
| Thinking budget control | User-specified multiples of 512 tokens (0 = no thinking, 512/1K/2K/4K/8K/16K recommended)github.com ↗ | verified |
| Inference backends | transformers ≥4.56.1, vLLM, SGLang; bfloat16 recommendedgithub.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AIME24 Open-source SOTA among models with thinking-budget control; performance improves with larger budgets on hard math. | 91.7%% | github.com ↗ | |
AIME25 Strong performance on the newer AIME split; second only to selected closed models in public comparisons. | 84.7%% | github.com ↗ | |
LiveCodeBench v6 Open-source SOTA on LiveCodeBench (Feb–May 2025 window); thinking budget yields measurable gains on harder problems. | 67.4%% | github.com ↗ | |
SWE-Bench Verified (OpenHands) Leading open-source result on agentic software engineering benchmark; strong issue-resolution capability. | 56.0%% | github.com ↗ | |
RULER (128K) Matches or exceeds closed models on long-context retrieval with native 512K training. | 94.6%% | github.com ↗ | |
MMLU-Pro Strong general-knowledge result; competitive with much larger models. | 82.7%% | github.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "bytedance/seed-oss-36b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"bytedance/seed-oss-36b","messages":[{"role":"user","content":"Explain the difference between GQA and MQA in one paragraph."}]}'ByteDance's Seed Team open-sourced the full Seed-OSS-36B family (Base w/ syn, Base-woSyn, Instruct) with 512K native context, thinking-budget control, and strong agentic benchmarks. The release emphasizes research-friendly variants and international (i18n) optimization.
Users can pass an integer multiple of 512 (e.g., 512, 1024, 2048) as the thinking budget. The model performs chain-of-thought reasoning up to that token limit, then produces a final answer. Setting budget=0 disables thinking entirely. The feature is exposed by some providers via an extra_body parameter; TheRouter forwards it when the upstream supports it.
It shows leading open-source results on SWE-Bench Verified and TAU1 benchmarks, but as a recent August 2025 release it has limited third-party production telemetry. Operators should run their own evaluation harness and consider keeping a fallback route to more mature models (DeepSeek-V3.2, Qwen3-235B) until broader adoption data is available.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | github.com ↗ | 2026-05-25 | verified |
| License | github.com ↗ | 2026-05-25 | verified |
| Architecture | github.com ↗ | 2026-05-25 | verified |
| Context window | github.com ↗ | 2026-05-25 | verified |
| Training data | github.com ↗ | 2026-05-25 | verified |
| Thinking budget control | github.com ↗ | 2026-05-25 | verified |
| Inference backends | github.com ↗ | 2026-05-25 | verified |
| AIME24 | github.com ↗ | 2026-05-25 | verified |
| AIME25 | github.com ↗ | 2026-05-25 | verified |
| LiveCodeBench v6 | github.com ↗ | 2026-05-25 | verified |
| SWE-Bench Verified (OpenHands) | github.com ↗ | 2026-05-25 | verified |
| RULER (128K) | github.com ↗ | 2026-05-25 | verified |
| MMLU-Pro | github.com ↗ | 2026-05-25 | verified |
| ByteDance Seed releases Seed-OSS-36B family under Apache 2.0 | github.com ↗ | 2026-05-25 | verified |
| How does the thinking budget feature work? | github.com ↗ | 2026-05-25 | to verify |
| Is Seed-OSS-36B suitable for production agent deployments? | github.com ↗ | 2026-05-25 | to verify |