Seed-Coder is a family of 8B parameter code LLMs (base, instruct, reasoning) from ByteDance Seed, designed to enhance coding capabilities by using LLMs to curate their own training data, minimizing human effort. It targets developers and researchers seeking powerful, lightweight, open-source code intelligence solutions.

How It Works

Seed-Coder employs a "model-centric" data pipeline, leveraging LLMs for data filtering and curation from sources like GitHub, commits, and web data. This approach aims to reduce manual effort in pretraining data construction while achieving state-of-the-art performance for its size.

Quick Start & Requirements

• Install/Run: Deployable via Hugging Face transformers or vLLM.
• Prerequisites: torch (bfloat16 recommended), transformers, vLLM for advanced deployment.
• Resources: 8B parameter models; vLLM supports multi-GPU and tensor parallelism for long contexts (up to 32K tokens).
• Links: Homepage, Hugging Face

Highlighted Details

• State-of-the-art performance among open-source models at the 8B scale across various coding tasks.
• Models include Base, Instruct (for user intent alignment), and Reasoning (RL-trained).
• Supports long context windows up to 64K tokens for the Reasoning model.
• Fully compatible with vLLM for efficient inference and distributed serving.

Maintenance & Community

• Developed by the ByteDance Seed Team, founded in 2023.
• Models are publicly available on Hugging Face.

Licensing & Compatibility

• MIT License. Permissive for commercial use and closed-source linking.

Limitations & Caveats

• Evaluation results for BigCodeBench were updated due to an inconsistent setting; users should refer to the latest reported benchmarks.