ByteDance Seed open-sources Seed-Coder

Background

A code model is a language model trained mainly on programming code. Developers use such models to complete code, write functions and explain existing programs. A model's parameter count, written as 8B for 8 billion, is the number of adjustable values it learns during training. Pre-training is the first large training stage, where the model learns from a broad body of data before any task-specific tuning.

What was released

On 8 May 2025 ByteDance Seed released Seed-Coder, a family of open code models with 8 billion parameters each.^[1]

The family has three variants. Base is the pre-trained model. Instruct is tuned to follow written requests, such as a request to write a function. Reasoning is tuned to work through a problem step by step before answering.^[1]

The context length, which is the amount of code and text the model can take in at one time, is 32K tokens for Base and Instruct, and 64K tokens for Reasoning.^[1]

How the data was prepared

The pre-training data was curated mainly by large language models rather than by hand-written rules. In other words, a language model was used to select and filter the code that the Seed-Coder models learn from.^[1]

Performance and license

ByteDance claimed state-of-the-art results among open models of its size.^[1] The models were released under the MIT license, a permissive license that lets others use, copy, modify and redistribute the code.^[1]

Earlier name

The project had previously been called Doubao-Coder.^[1]

Code models are usually judged on tasks such as completing a function from its description or fixing a bug in a short program. Code has strict syntax and can be executed, so a model's output is easier to check objectively than much other generated text.

Significance

The three variants cover completion, instruction following and step-by-step reasoning within one family.