Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

Seed-OSS releases three 36B model checkpoints

ByteDance Seed published two base variants and one instruction-tuned variant of Seed-OSS-36B. The model card describes training for contexts up to 512K tokens. [1][2]

Model access
Downloadable 36B Instruct weights and two 36B base checkpoints. [1][2]
License
Apache License 2.0 for the published Seed-OSS-36B checkpoints. [1][2]

What the model does

On August 20, 2025, ByteDance Seed released the checkpoints under Apache 2.0. One base checkpoint leaves synthetic instruction data out of pretraining so researchers can examine that choice. [1][2]

Why it matters

The two base variants let researchers study the effect of one training-data choice while keeping model size fixed. The instruction-tuned variant gives application teams a separate starting point. [1][2]

Who can use it

Developers evaluating longer inputs can use the Instruct model. Teams studying later training can choose between the two documented base checkpoints. [1][2]

Key dates

  1. Seed-OSS-36B variants released [1]

Ways to use it

Evaluate responses against relevant passages distributed through a long document. Test the input lengths the application will actually send and record the serving setup. [2]

Model limitations

The documented context length describes what the model was trained to accept. It does not ensure accurate use of every passage or make long requests practical on all hardware. [1][2]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source ByteDance Seed’s dated Seed-OSS-36B model card
  2. Primary source ByteDance-Seed/Seed-OSS-36B-Instruct

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked