Model

Seed-OSS-36B

OPEN

ByteDance · Seed family · released Aug 20, 2025

Apache-2.0 open release with native 512K context and a tunable thinking-length budget; trained on only 12T tokens.

ReasoningCodingVisionFunction callingTool useAgentic

1576.9

Elo · rank #122

Parameters

36B

Active params

36B (dense)

Context

512K tokens

Architecture

Dense Transformer (RoPE, GQA, RMSNorm, SwiGLU)

License

Apache-2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath91.7%#40

best: GPT-5.2 · 100.0%

ARC-AGI-2Reasoning1.4%#29

best: GPT-6 Astra · 95.0%

GPQA DiamondReasoning71.4%#125

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning10.1%#97

best: Claude Opus 5 · 64.7%

IFEvalReasoning85.8%#53

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding67.4%#78

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MMLU-ProKnowledge82.7%#43

best: Claude Fable 5 · 91.5%

MMLUKnowledge87.4%#34

best: OpenAI o3 · 92.9%

MMMLUKnowledge78.4%#31

best: Gemini 3.1 Pro · 92.6%

SimpleQAKnowledge9.7%#38

best: GPT-4.5 · 62.5%

SuperGPQAReasoning55.7%#17

best: Qwen3.7-Max · 73.6%

SWE-bench VerifiedCoding56.0%#89

best: Claude Opus 5 · 96.0%

tau-benchAgents70.4%#6

best: Claude Opus 4.1 · 82.4%

Run it locally

VRAM @ Q4

21.8 GB

VRAM @ FP16

72 GB

Fits on (Q4)

RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4 · GGUF Q8

Fine-tune it

Permissive

QLoRA24.7 GB1× RTX 5090 32GB

LoRA76.9 GB1× A100 80GB

Full fine-tune578.0 GB4× B200 192GB

QLoRA SFT on ~10k samples ≈ $17.55 (1× RTX 5090 32GB)

Plan a fine-tune →

Seed family

Elo progression across releases

Seed1.5-ThinkingApr 20251598.4 Seed-OSS-36BAug 20251576.9

Resources

HFByteDance-Seed/Seed-OSS-36B-Instruct GHByteDance-Seed/seed-oss

API price weights · each benchmark row carries its own source badge (see methodology)