Open-source 8B code-LLM family (previously branded "Doubao-Coder"). Three released variants: Base (32K), Instruct (32K), and Reasoning (64K, RL-trained). The technical contribution is "LLMs can effectively curate code training data by themselves" — a model-centric data-curation pipeline that beats rule-based filters and produces SOTA results among 8B open-source code models on generation, completion, editing, reasoning, and SWE tasks.

Successor in branding (and pretraining recipe) to the closed Doubao Seed Code flagship. MIT license.

Model Details

Architecture DENSE
Parameters 8B
Context window 65,536
License MIT

Variants

Name Parameters Notes
Seed-Coder-8B-Base 8B 32K context
Seed-Coder-8B-Instruct 8B 32K context
Seed-Coder-8B-Reasoning 8B 64K context, RL-trained

Paper

open-weightcodingreasoning

Related