AI Research Atlas

Seed-OSS-36B

ByteDance Seed · 20 August 2025

ByteDance's first open LLM family is a 36B dense model with 512K native context, Apache-2.0 licence and an adjustable thinking budget from 0 to 16K tokens.

Trained on 12T tokens with GQA; users set a thinking budget (512 to 16K tokens) and the model self-reflects at intervals. Reports 91.7 AIME24, 67.4 LiveCodeBench v6, 56 SWE-bench Verified, 94.6 RULER at 128K.

Date
Wednesday, 20 August 2025
Lab
ByteDance Seed
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
AIME24 (Instruct)91.7company
LiveCodeBench v667.4company
SWE-bench Verified56company
Native context length512K tokenscompany

A Base-woSyn variant is also published. Benchmarks are ByteDance-reported.

Sources

  1. huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct
  2. github.com/ByteDance-Seed/seed-oss

This record was checked against its sources on 6 October 2026. How we check