Seed-Coder Models by ByteDance

ByteDanceMITOpen source

3 models2025Up to 66k ctx

Last refreshed 2026-06-21. Next refresh: weekly.

Details

ResearcherByteDance

LicenseMITOSI-approved

Commercial useCommercial use: permitted

Models3

Released2025

Max context66k

Capabilities

Reasoning1 of 3 models

Links

Website HuggingFace

About

Seed-Coder is a family of lightweight open-source code LLMs developed by ByteDance Seed, comprising base, instruct, and reasoning models trained on 6 trillion code-centric tokens using a model-centric data curation approach. Previously known internally as Doubao-Coder. All models are 8B parameters and available under MIT license on Hugging Face.

Decision facts

Best fit: codereasoningcoding
Capability starting point: Seed-Coder 8B Reasoning with 66k context and reasoning
Lowest tracked input: Not tracked
Closest related family: BAGEL

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

3 in view

Seed-Coder 8B BaseCurrent

Use when the workload needs code, 33k context, and 8B parameters.

2025-05code33k context8B parameters

Seed-Coder 8B InstructCurrent

Use when the workload needs code, 33k context, and 8B parameters.

2025-05code33k context8B parameters

Seed-Coder 8B ReasoningCurrent

Use when the workload needs code, 66k context, and 8B parameters.

2025-05code66k context8B parameters

Current Seed-Coder variants with use-when guidance and lifecycle status
Model	Use when	Released	Signals	Status
Seed-Coder 8B Base	Use when the workload needs code, 33k context, and 8B parameters.	2025-05	code33k context8B parameters	Current
Seed-Coder 8B Instruct	Use when the workload needs code, 33k context, and 8B parameters.	2025-05	code33k context8B parameters	Current
Seed-Coder 8B Reasoning	Use when the workload needs code, 66k context, and 8B parameters.	2025-05	code66k context8B parameters	Current